Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Development of a Residential Road Collision Warning Service Based on Risk Assessment

Development of a Residential Road Collision Warning Service Based on Risk Assessment Hindawi Journal of Advanced Transportation Volume 2023, Article ID 7496377, 16 pages https://doi.org/10.1155/2023/7496377 Research Article Development of a Residential Road Collision Warning Service Based on Risk Assessment 1 2 3 4 Aya Selmoune , Jeongin Yun , Myoungkook Seo , Hyeokhyeon Kwon , 5 2 Changhee Lee , and Jinwoo Lee Jiangsu Key Laboratory of Urban ITS, Southeast University, Nanjing 210000, China Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon 34051, Republic of Korea Smart Engineering Laboratory, Korea Construction Equipment Technology Institute, Gunsan 54004, Republic of Korea Research and Development Center, Pintel Incorporated, Seoul 06729, Republic of Korea Transportation Policy Division, Daejeon Metropolitan City Hall, Daejeon 35242, Republic of Korea Correspondence should be addressed to Jinwoo Lee; lee.jinwoo@kaist.ac.kr Received 9 August 2022; Revised 4 February 2023; Accepted 11 February 2023; Published 16 March 2023 Academic Editor: Jose E. Naranjo Copyright © 2023 Aya Selmoune et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Pedestrians are more likely to be seriously injured in vehicle collisions. In fact, multiple collisions between vehicles and pedestrians occur on residential roads that lack street-to-sidewalk dividers and have numerous blind spots. Traditional trafc safety features and equipment, such as speed bumps and trafc signs, are not always sufcient to prevent pedestrian accidents on such residential roads. Terefore, we suggest a collision risk warning service for residential roads as a solution to this issue. We use CCTVs with computer vision techniques and radar to accurately detect objects in real-time and to trace their trajectories. In addition, we employ a time-to-collision-based method to identify dangerous situations. Te service warns drivers and pedestrians about hazardous situations using a light-emitting diode sign board. We applied our service to three diferent roads on a university campus in Seoul, Korea, and then conducted a user survey to evaluate the service. In summary, more than 90% of respondents stated that the service was necessary for these specifc locations, and 76.9% noted that the service signifcantly contributed to trafc safety on the campus. Tis implies that the proposed service improved trafc safety and can be applied to various locations on residential roads. [6]. Pedestrians are the most likely to be seriously injured in 1.Introduction vehicle collisions. Traditional trafc safety features and Approximately 1.3 million people die annually because of equipment, such as speed bumps and trafc signs, are not trafc accidents [1]. Some governments and agencies in always sufcient for preventing pedestrian accidents in blind many countries have tried to reduce trafc accidents by spots on residential roads. Particularly when pedestrians abruptly exit from parked vehicles on roads, drivers are implementing safety education and policies such as pro- moting trafc rules and enforcing speed limits [2]. As a result unable to respond appropriately, and trafc accidents are of these eforts, trafc fatalities in most developed countries highly possible. in the OECD have decreased substantially. For example, Several technologies have been developed to prevent Korea reduced road fatalities by 26.4% from 2017 to 2020 [3]. vehicle-pedestrian collisions. Tey are based on algorithms However, safety issues of pedestrians remain a concern [4]. that identify objects, predict their trajectories, and determine Pedestrian fatalities in Korea accounted for 35% of total whether or not a collision risk exists. Te algorithms can be fatalities [5]. More than half of pedestrian fatalities occur on divided into two categories depending on how the collision residential roads without separation of streets and sidewalks risk is determined. First, some algorithms employ surrogate 2 Journal of Advanced Transportation drivers. Few services considered a pedestrian perspective. safety measures (SSMs) to recognize the presence of po- tentially dangerous situations based on whether predicted One study developed a system that recognized dangerous situations and provided information to pedestrians via their trajectories of objects overlap. Using microscopic trafc characteristics such as vehicle speed, acceleration, time smartphones [20]. However, it was inaccurate and in- headway, and space headway, an SSM method assesses the efective in that object detection was conducted only by collision risk of particular trafc scenarios [7]. SSMs, such as cameras on smartphones. In addition, few studies evaluated time-to-collision (TTC) and post encroachment time (PET), the efects of proposed algorithms in the feld. Most algo- have been widely used to evaluate trafc safety performance rithms were evaluated based on simulations or feld pro- and identify potential accident risks [8–13]. One study as- totype tests, and accuracy was only verifed through a confusion matrix. sumed a connected environment in which pedestrians and vehicles shared real-time location information using IoT In the present study, we propose a safety service framework that provides risk information to both vehicles devices. According to object locations, velocity, relative distance, angle, and TTC, dangerous situations were de- and pedestrians. Te proposed framework utilizes RSE such as CCTVs and radar to detect objects using a deep learning termined [8]. In another study, an algorithm was developed using onboard cameras in vehicles. Potential collision areas method. Ten, the algorithm uses SSMs to identify whether were defned by the minimum TTC from the predicted the current situation is dangerous. If the situation is unsafe, movements of ego vehicles and pedestrians [9]. In addition, a light-emitting diode (LED) sign board gives warning in- in a connected vehicle environment, a crash warning system formation to both vehicles and pedestrians to avoid a po- was developed for bike lane areas. PET was used to identify tential collision. Tus, the service alerts drivers and potential areas of interaction between vehicles and bicycles pedestrians at the same time. To evaluate the safety efects of the proposed service, we implemented and operated it on- [10]. Most algorithms were verifed as simulation-based or autonomous platforms. Te second set of algorithms pre- site. We conducted a survey to investigate user satisfaction. Te remainder of this paper is structured as follows: the dicts risk situations using deep learning methods [11–13]. After an algorithm is trained using prior data labeled by service description section presents the overall framework. Te application and evaluation section introduces the study SSMs as risk situations, it predicts whether a given situation is dangerous. Te gated recurrent unit method was used to site and presents the evaluation. In the last section, we predict collision risk at a signalized intersection [11]. Sim- summarize this study and discuss possible future research ilarly, long-short term memory (LSTM) was used to predict directions. risk situations [12]. In some cases, deep learning methods were used for trajectory estimation to predict risk situations. 2. Service Description One study proposed a collision risk area estimation system at unsignalized crosswalks. Te system used LSTM to predict We propose a collision risk warning service procedure, as object trajectories and then conducted statistical inferencing depicted in Figure 1. Tis service is a proactive counter- to predict collision risk areas [13]. measure against vehicle-vehicle or vehicle-pedestrian col- As soon as a potentially hazardous situation is identifed, lisions. Tere are four steps: Step 1 object detection through various warning services are provided. Tis warning in- CCTV and radar; Step 2 trajectory prediction of detected formation can be divided into three categories. First, in- objects; Step 3 collision risk identifcation based on predicted formation is provided by vehicles. Augmented reality (AR) trajectories; and Step 4 collision risk warning, if any. Here, on the heads-up display in vehicles was employed to display the current time is t , and the previous point one time step warning information. In addition to AR, an audio warning before and the future point n time steps after are denoted by was immediately provided [10]. Several active pedestrian t and t . −1 n collision avoidance systems did not give alerts but instead automatically controlled the vehicles [14, 15]. Te second method is to provide information to vehicles from roadside 2.1. Object Detection. We use CCTV and radar equipment to equipment (RSE). For example, amber fashing lights were detect vehicles and pedestrians. One of the detecting algo- activated when pedestrians were approaching or crossing rithms is you only look once (YOLO) [21], which has been crosswalks so that the drivers could perceive them [16]. Te used in various felds for real-time detection. We employ third method is to use infrastructure-to-vehicle (I2V), ve- a YOLO v5-based algorithm. YOLO v5 is faster and more hicle-to-pedestrian (V2P), and vehicle-to-everything (V2X) accurate than its previous versions [22]. To account for the communication. One study utilized I2V communication to characteristics of residential roads, we need residential-road- give warning information to vehicles from RSE [17]. Several specifc training datasets, which are distinguished from studies developed V2P and V2X communication-based general road datasets. Terefore, we used 150 hours of video warning services in Wi-Fi environments [18, 19]. How- data from CCTV cameras installed on residential roads in ever, in the current state, the communication-based safety Guro-gu, Seoul. We trained for various environments such warning method has problems regarding latency and as lighting and weather conditions as well as situations stability. involving numerous objects such as pedestrians carrying Most systems were developed from the perspective of umbrellas, as shown in Figure 2. With the trained model, vehicles. Based on cameras or radar sensors in vehicles and objects can be accurately identifed in real-time as pedes- CCTVs in RSE, warning information was provided to trians, motorcycles, bicycles, vehicles, and personal mobility Journal of Advanced Transportation 3 Step 1. Object detection Transmitted pulse Reactive pulse Radar CCTV Step 2. Trajectory prediction t t t t t t –1 0 n –1 0 n Step 3. Collision risk identifcation Dangerous Not dangerous Step 4. Collision risk warning Figure 1: Overall structure of collision risk warning service. devices, even under severe lighting and weather conditions. in Table 1. At the 50% level of Intersection over Union, In Figure 3, the training results are shown with an example defned as the degree of overlap between ground truth and site at two diferent time points compared to the identif- prediction regions [23, 24], the detection rate for pedes- cation without the training. Te overall accuracy is presented trians, motorcycles, bicycles, vehicles, and personal mobility 4 Journal of Advanced Transportation Table 1: Post-training object detection accuracy. Pedestrian Motorcycle Bicycle Vehicle Personal mobility device Proposed algorithm 2573/2590 2213/2216 2994/2995 341/343 934/942 (Detection rate) (99.34%) (99.86%) (99.97%) (99.42%) (99.15%) Number of accurately detected objects/total number of objects. Journal of Advanced Transportation 5 devices was higher than 99%. In addition, radar is employed Multiple object tracking accuracy (MOTA) is used to to complement CCTVs. Tey provide precise locations and evaluate the accuracy of object-tracking algorithms [32]. MOTA is the most prevalent indicator used to measure speeds, which are difcult to collect with CCTVs. With these two complementary devices, accurate and precise real-time a tracker’s performance. Its value may be determined using object detection is achieved. equation (3). IDSW + FN + FP 2.2. Trajectory Prediction. Future trajectories of objects are (3) MOTA � 1 − , GT predicted based on their previous coordinates that we can track. In this study, perspective transformation and Kalman flter are where ground truth (GT) is the total number of ground truth used for tracking objects. Coordinates of objects detected by objects, identity switching (IDSW) represents the number of CCTVs are transformed into overhead perspectives to measure ID switches in the video stream, false negative (FN) indicates exact locations. We employ the perspective matrix in Open CV a missed detection, and false positive (FP) means an in- to convert the coordinates from the videos to overhead co- accurate detection. Based on MOTA, we observe that the ordinates [25, 26]. Kalman flter involves repeating the pre- proposed algorithm results in higher accuracy than Deep- diction step and correction step of trajectories [27]. In the SORT. Mostly tracked targets (MT) and mostly lost targets prediction step, the next position of the object in the current time (ML) are the number of tracked and lost objects, re- is estimated based on the information collected about the object spectively. Te proposed algorithm has higher and lower already being tracked, as in equation (1). values of MT and ML, respectively, than DeepSORT, which x � Ax + w , (1) are desirable. Furthermore, the proposed object-tracking k k−1 k−1 algorithm has a higher FPS than DeepSORT since we do where x is the state vector representing the object’s dynamic k not use the computationally burdening Hungarian matching behavior at a discrete time index k; A is the transition matrix algorithm. In summary, the proposed tracking algorithm at time index k − 1 to k; and vector w is the noise fol- k−1 outperforms DeepSORT. lowing normal probability distribution N(0, Q) with zero Ten, based on the tracking data, we predict vehicle and mean and covariance matrix Q. In the correction step, the pedestrian trajectories. First, we classify straight and curved previously-predicted position is compared to the position trajectories based on whether the angle of the previous measured by CCTVs. To modify the object position, a weight trajectories is smaller than the angle that we predetermine, called Kalman gain K is used, which indicates the ratio of k θ > 0 (unit: radians). We estimate a vehicle’s tendency to set the error of the predicted object position to the error of the move based on the angle diference between previous points. object position measured by the object detection algorithm. Ten, we refect this tendency in the trajectory prediction. If K has a range from zero to one, and it is infuenced by more we set the time index k to zero for the current time, the accurate values between the predicted position and the ∗ current position is P at x , and the previous positions at 0 k�0 measured position, as stated in equation (2). two and one time steps before are P and P , respectively, −2 −1 − − ∗ ∗ x 􏽢 � x 􏽢 + K z − Hx 􏽢 􏼁 , (2) i.e., the coordinates of P and P are x and x . Te k k k k k −2 −1 k�−2 k�−1 �������→ �����→ known angle between P P and P P is θ (unit: ra- − −2 −1 −1 0 −1 􏽢 􏽢 where x is a posteriori estimated state; x is a priori es- k k dians). Te tendency angle θ (unit: radians) is calculated in timate; z is the observed measurement; and H is the equation (4). measurement matrix at time k − 1  to  k. After the repeated execution, we update K and fnd the optimal state (x ) that k 0 minimizes the error between the estimated state and the θ � 􏽘 θ ω , (4) 0 k k measured state [28]. past k�−T Te object-tracking algorithm proposed in this where ω is a weight factor that considers the angle error, s.t., study was compared to DeepSORT, which is a deep past ω � 1, and T is the time window length of the past learning-based method for tracking objects [29, 30]. k�−T k data, i.e., we consider the tracking data at DeepSORT consists of four key components: detection, past past k ∈ −T , −T + 1, . . . , −1, 0 . 􏼈 􏼉 estimation, data association, and generation and de- We consider a short prediction period between k � 0 and letion of tracking objects. In DeepSORT, Kalman flter future future past k � T , where T ≪ T . Tus, it can be reasonable is used in the estimation stage, and a Hungarian to assume that a vehicle with |θ | < θ keeps moving straight 0 set matching algorithm is employed in the data associa- future future during T . At future time point n ∈ 􏼈1, . . . , T − 1􏼉, tion stage [30]. Te major diference between Deep- the center location of the straight moving vehicle, denoted SORT and the proposed algorithm is twofold. First, we �����→ P , is on the straight line extended from P P and the n −1 0 advanced the Kalman flter algorithm. Second, due to ����→ the Hungarian algorithm’s prohibitive computational 􏽢 􏽢 distance from P to P , |P P |, is the multiplication of the 0 n 0 n cost, we developed an original matching algorithm average vehicle speed and the time diference between t and instead of using the Hungarian algorithm. For com- t . If |θ | ≥ θ , the center of a vehicle predicted at time point n 0 set parison, we used the Oxford Town Centre dataset, 􏽢 􏽢 n, P , is found based on P , θ , and the average vehicle n n−1 n−1 which is commonly employed to assess object-tracking speed. Particularly, P is calculated using P and θ . Similar 1 0 0 performance [31]. Te comparison results are shown in n −1 􏽢 􏽢 to equation (4), θ is defned as 􏽐 θ ω +􏽐 θ ω Table 2. n k�0 k k−n k�n−T k k−n 6 Journal of Advanced Transportation (a) (b) (c) (d) (e) (f) Figure 2: Various environments for training model. Te model was trained in various lighting conditions, including clear (a, b, c), snowy (d), and rainy weather (e, f). future range that the pedestrians reach varies. Te estimated for all n ∈ 􏼈1, . . . , T 􏼉. Te spatial range of a vehicle, major and minor axes of the ellipse are determined using predicted at n, is defned to have its center at P , and its equation (5), where the parameters were tuned based on boundary is determined based on the actual vehicle size. the actual data. Compared to vehicles, pedestrians have relatively uncertain and inconsistent movement characteristics. major ⎧ ⎪ 1.6 d − 0.89, 􏼐e ≥ 1.0􏼑, ⎪ n n Tus, we use an elliptical trajectory prediction approach major e � to account for stochastic pedestrian trajectories, and we ⎪ major 1.0, 􏼐e < 1.0􏼑, consider the ellipse as the future spatial range of a pe- destrian’s location [33]. We estimate the moving di- rection of pedestrians based on the previous directions in ⎧ ⎪ 2.0, d ≥ 3.0 􏼁 , (5) the same method for vehicles, as described in equation minor e � (4). We estimate the major and minor axes of an ellipse ⎪ 1 using actual pedestrian path data collected from CCTVs ⎪ ⎩ d + 1.0, d < 3.0 , n n on residential roads [34]. Estimation results depend on the time point t for all at which we predict from the k�n 􏼌 􏼌 􏼌 􏼌 future 􏽢 􏼌 􏼌 present time, n > 0. We can fnd P of the center of an 􏼌 􏼌 d � P P , ∀n ∈ 1, . . . , T , 􏽮 􏽯 n 􏼌 􏼌 n 0 n ellipse for a pedestrian in a similar way to fnding that for major a vehicle with |θ | ≥ θ . Depending on the moving dis- where e (unit: meters) is the major axis of the ellipse at 0 set minor tance d (unit: meters) from the current time point t to future time point n, and e (unit: meters) is the minor n 0 n the future point t , defned as |P P |, the possible spatial axis of the ellipse at future time point n. n 0 n Journal of Advanced Transportation 7 (a) (b) Figure 3: Detection performance improvement: (a) before and (b) after additional training. After additional training, the model detects a pedestrian carrying umbrella and distinguishes a personal mobility device from a pedestrian. Table 2: Object-tracking performance comparison. MOTA IDSW FN FP MT ML GT FPS (%) DeepSORT 79.90 1024 5667 2753 114 2 46985 68.65 Proposed tracking 81.70 743 3590 4258 136 1 46945 333.79 algorithm Te predicted trajectories are graphically illustrated in coordinate of P . Te MAE was calculated to be between 0.09 Figure 4. and 0.65 meters. Te accuracy is lowered as the curvature We tested the accuracy of the trajectory prediction model and speed increase. using example trajectories of three pedestrians and two vehicles, as depicted in Figure 5. Te length of each time period is one 2.3. Collision Risk Identifcation. Once future trajectories of frame, and for each prediction, we plotted the center of a pe- objects intersect, we get one intersecting point destrian’s ellipse or a vehicle at n � 3 (after three seconds from (q 􏽢 􏽢 , s ) of two objects a and b, and two collision time each prediction time point). Te prediction trial is indexed by i intersect intersect (CT) to reach the intersecting point from each object at the and the total number of trials is I. We conducted 857, 396, 899, current location. q 􏽢 and 􏽢 s are calculated by equation 330, and 324 trials for pedestrian #1, pedestrian #2, pedestrian intersect intersect (7) [35]. #3, vehicle #1, and vehicle #2, respectively. Te test results are b a presented in Figure 6 and Table 3. Te unit of the graphs in b a b a 􏽢 􏽢 􏼐s − s 􏼑 − 􏼒q tan θ − q tan θ 􏼓 0 0 0 0 0 0 Figure 6 is in pixels, and the horizontal and vertical lengths of q 􏽢 � , intersect a b a pixel are 0.09 meters and 0.11 meters, respectively. For the 􏽢 􏽢 tan θ − tan θ 0 0 accuracy measure, we use mean absolute error (MAE) calculated (7) by equation (6). b a b a b a 􏽢 􏽢 􏼐q − q 􏼑 − 􏼒s cot θ − s cot θ 􏼓 􏼌 􏼌 0 0 0 0 0 0 􏼌 􏼌 I 􏼌 i i 􏼌 􏼌 􏼌 x 􏽢 − x 􏽢 s � , 􏼌 􏼌 i�1 3 3 intersect a b (6) 􏽢 􏽢 MAE � , cot θ − cot θ 0 0 i i where q 􏽢 and 􏽢 s represent the longitudinal and where x 􏽢 and x are the predicted and the actual locations of intersect intersect 3 3 latitudinal coordinates of the intersecting point, respectively; the object three seconds later than the current time point of a a b b th i (q , s ) and (q , s ) represent the current coordinates of the i prediction trial, respectively. Specifcally, x 􏽢 is the 0 0 0 0 3 8 Journal of Advanced Transportation |θ | < θ 0 set –1 P ⌃ –1 P θ n –1 –2 Straight line –1 1 Vehicle’s spa spat ti ial ra al ran ng ge e |θ | ≥ θ 0 set θ P –1 P –1 –1 –2 Curved line P edestrian’s spatial range (a) (b) Figure 4: Trajectory prediction method: (a) trajectory classifcation—straight and curved trajectories are classifed based on whether the angle of the previous trajectories is smaller than the angle predetermined. (b) Trajectory predictions—vehicle and pedestrian trajectories are predicted as rectangles and ellipses, considering the actual vehicle size and stochastic pedestrian trajectories, respectively. Pedestrian trajectory Vehicle trajectory Figure 5: Test scenario description. Trajectories of three pedestrians and two vehicles were tested. a b a 􏽢 􏽢 􏽢 objects a and b, respectively; and tan θ , tan θ , cot θ , and considering the diference between the center point and the 0 0 0 cot θ represent the current tangent and cotangent values of spatial range boundary of each object. objects a and b. With q 􏽢 , 􏽢 s , and the objects’ speeds intersect intersect If the spatial ranges of two objects successively and directions, we can calculate CTs for both objects, overlap at more than three intervals, the time interval a b 􏽣 􏽣 CT   and  CT [36]. We compare the two CTs, and the when the two spatial ranges frst overlap is defned as the a b predicted TTC. Tis indicator assumes that the involved 􏽣 􏽣 smaller one, CT ( ≡ min CT , CT , unit: seconds), is 􏼚 􏼛 min objects do not recognize the risk and there is no urgent used in determining whether the spatial ranges overlap in maneuver to avoid it in a following short period of time. increments of 0.25 seconds starting from one second earlier We compare the predicted TTC with a TTC threshold to than CT (i.e., CT − 1 + 0.25δ  for  δ � 0, 1, 2, . . .) min min identify whether a collision risk exists [37–39]. If it is Journal of Advanced Transportation 9 (unit: pixels) Pedestrian #1 Pedestrian #2 Pedestrian #3 620 640 660 680 700 0 500 1000 1500 620 640 660 680 700 0 300 Expected route Expected route Expected route Real route Real route Real route Vehicle #1 Vehicle #2 620 640 660 680 700 0 200 400 600 800 1000 0 370 900 420 Expected route Expected route Real route Real route Figure 6: Trajectory prediction test results. Table 3: Trajectory prediction accuracy. Pedestrian Vehicle #1 #2 #3 #1 #2 MAE 0.09 0.39 0.18 0.54 0.65 (unit: meters) Current state Future state: After CT – 1 + 0.25δ seconds (δ = 0,1,2,...) min CT seconds ... CT seconds Object a ... CT – 1.00 CT – 0.75 CT – 0.50 CT – 0.25 CT Object b min min min min min a b here, predicted TTC = CT – 0.50 (units: seconds) min CT ≡ min {CT , CT } min Spatial range overlap (a) (b) a b 􏽣 􏽣 Figure 7: Determination of risk existence and prediction of time-to-collision (TTC): (a) CT is determined based on CT and CT . (b) If min two spatial ranges overlap successively more than three intervals, the predicted TTC is calculated as the frst overlap, and if the predicted TTC is smaller than the TTC threshold, the collision risk is regarded to exist. 10 Journal of Advanced Transportation smaller than the TTC threshold, the collision risk is status and the service on the site, as shown in Figure 10. regarded to exist. Te entire process of the TTC calcu- 32.5% of respondents reported that they had encountered unsafe situations on KAIST Seoul Campus, with 44.9% in A, lation is depicted in Figure 7. We assume that the TTC threshold value is the summation 22.4% in B, and 10.3% in C. 30.6% of respondents who of the perception reaction time, the margin time for an LED sign, experienced accident risk indicated that those risks had and the vehicle stopping time. In this study, a fxed TTC occurred at night. Moreover, 26.5% were using a smart- threshold of four seconds is used to account for a safety margin phone, and 22.4% were wearing earphones (or headphones) to some extent. We consider the perception reaction time to be when the incidents happened. Regarding campus trafc 1.5 seconds [40], the LED sign margin time to be 1 second, and safety, 78.8% of respondents indicated that it should be the vehicle stopping time to be 1.5 seconds. improved for four reasons, as shown in Figure 11. First, there is no separation between streets for vehicles and sidewalks for pedestrians on campus roads (16.8%). Second, the road 2.4. Collision Risk Warning. If the situation is judged to be widths are narrow (15.1%). Tird, insufcient guiding signs dangerous, drivers and pedestrians are presented with LED sign on one-way roads frequently lead to wrong-way driving information. Tis service delivers warning information on the (7.6%). Fourth, there are multiple blind spots due to parked roadside for vehicles and pedestrians, as opposed to prior vehicles and buildings (5.9%). Tese factors are consistent systems that provided risk information only to vehicles. Te with the safety problems of residential roads in other regions warning information is presented in Figure 8. of Korea [6]. Ten, after service installation and operation on KAIST Seoul Campus, 76.9% of total subjects noted that the 3.Application and Evaluation service contributed to trafc safety on the campus from the 3.1. Application. We applied the proposed service on KAIST results of the Likert scale, as shown in Table 4. Tey stated Seoul Campus in Korea. Te service was provided in three that the service could prevent collision risk in the blind spots situations: illegal roadside parking, unprotected left turn, by providing warnings. Specifcally, they mentioned that and wrong-way driving. Figure 9 provides a description of LED sign boards made signs instantly recognizable, even at the application site and each situation. In addition, we night, compared to convex mirrors. In addition, 78.6%, evaluated the proposed service to analyze its impact on safety 77.6%, and 85.1% of respondents who were aware of the by conducting an on-site survey. service in Locations A, B, and C, respectively, believed that roads were safer after service operation. In addition, 92.9%, 91.8%, and 94.6% of respondents who knew the service in 3.2. Survey Design. After the service application, we analyzed Locations A, B, and C indicated that the service was nec- responses to the service in terms of safety efects. We collected essary for campus trafc safety, as presented in Figure 12. data through in-person interviews. Te survey questionnaire Second, we conducted a chi-square test to determine consisted of four sections. First, we inquired about demographic whether respondents’ perceptions of the service in operation characteristics, including gender, age, mobility impairment, and at three diferent locations difered signifcantly. A chi- current modes of transportation. Second, we asked whether square test is a nonparametric test to analyze the in- accidents or accident hazards had occurred at the site. If so, dependence or diference across a group among nominal respondents were questioned as to whether they were using variables [41]. We used the chi-square test of homogeneity to a smartphone or headphones and about locations of incidents. compare the proportions of service perception among Furthermore, we inquired whether respondents thought campus groups at three locations for signifcant diferences. To trafc safety should be improved. Tird, for each location, we conduct the homogeneity test, samples of the test groups solicited feedback on the installed and operating service, such as must be distinct [42]. For this, three diferent groups of preference or level of satisfaction. We frst inquired if re- respondents were questioned about their opinions of the spondents were aware that the service was operational. If they service in the three designated locations (i.e., Locations A, B, were familiar with the service, they were asked about its safety and C). Each location had a diferent number of individuals efects following the operation and its requirements at each who were aware that the service was operational. In Location location. Finally, the contribution of the service to safety on C, there were more individuals who were unaware that the KAIST Seoul Campus was assessed using a fve-point service was operational than those who were aware, as Likert scale. shown in Figure 12. Terefore, we were able to use the Te respondents to the questionnaire were people who homogeneity test. Te formula for the test statistics, χ commute to KAIST Seoul Campus. Te survey was conducted -value, is mathematically expressed as equation (8). in July 2022. A total of 151 responses were collected. Te L J majority of respondents were campus members, such as stu- 􏼐O − E 􏼑 l,j l,j χ � 􏽘 􏽘 , (8) dents, professors, and employees, while some were local resi- l,j l�1 j�1 dents and travelers who were passing by KAIST Seoul Campus. Sample descriptions are summarized in Table 4. where O is observed frequency, E is expected frequency, l l,j l,j represents the location index, j refers to the category of 3.3. Survey Analysis. We analyzed the collected data, in- response (e.g., yes or no), and L and J are the number of cluding descriptive statistical analysis and the chi-square locations and the number of categories, respectively. Te test. First, we investigated the perception of the trafc safety degree of freedom (df) was found to be two, using the Journal of Advanced Transportation 11 Figure 8: Collision risk warning as delivered by LED sign board. Tree diferent images are displayed sequentially until a risk situation is over. Korea Advanced Institute of Science and Technology (KAIST) Seoul Campus Illegal roadside parking Unprotected lef turn Wrong-way driving (Location A) (Location B) (Location C) Parked vehicle LED sign board Driving vehicle One-way Driving vehicle (Wrong-way) Pedestrian Figure 9: Site description: the proposed service was applied to three diferent locations on KAIST Seoul Campus. Each location was chosen for a single specifc risk circumstance: Location A for illegal roadside parking; Location B for unprotected left turn; and Location C for wrong-way driving. 12 Journal of Advanced Transportation Table 4: Sample description. Sample attributes Number of samples % Male 114 75.5 Gender Female 37 24.5 20’s 78 51.7 30’s 41 27.2 Age 40’s 15 9.9 50’s 12 7.9 Over 60 5 3.3 Yes 1 0.7 Mobility handicapped No 150 99.3 On foot 117 77.5 Vehicle 26 17.2 Transportation mode Bicycle 5 3.3 Personal mobility 3 2.0 Less than six months 52 34.4 Working or visiting period Six months∼one year 26 17.2 More than one year 73 48.4 Yes 49 32.5 Accident risk experience No 102 67.5 Necessary 119 78.8 Necessity to improve trafc safety on KAIST Seoul Campus Not necessary 32 21.2 (Strongly) did not contribute 2 1.3 Did not contribute 5 3.3 Contribution of service to trafc safety Neutral 28 18.5 Contributed 99 65.6 (Strongly) contributed 17 11.3 <Transportation mode> (63.3%) <Gender> <Using a smartphone> (28.6%) 34 13 (6.1%) (26.5%) (69.4%) (2.0%) On foot Vehicle <Collision risk experience> (30.6%) (73.5%) Bicycle Personal mobility <Location> Male Use a smartphone Female Not use a smartphone 49 11 (67.5%) (32.5%) (22.4%) <Age> <Wearing earphones> (44.9%) (10.3%) (28.6%) (22.4%) (14.3%) (22.4%) (38.8%) A C 38 Yes 7 (77.6%) (14.3%) B Other No <Time> (4.0%) Wear earphones 19~29 50~59 (69.4%) Not wear earphones 30~39 Over 60 40~49 (30.6%) Daytime Nighttime Figure 10: Responses to questions about collision risk. Respondents who experienced risk were asked seven diferent questions. Journal of Advanced Transportation 13 <Need for safety improvement> (15.1%) (7.6%) <Reasons> No separation between streets and sidewalks 32 119 (5.9%) (16.8%) Narrow road widths (21.2%) (78.8%) High frequency of the wrong-way driving (0%) Numerous blind spots Steep slope No response Yes No Figure 11: Responses to question about need for safety improvement. <Is it safer than before at Location A?> <Is it safer than before at Location B?> (78.6%) (77.6%) <Awareness of service at Location A> <Awareness of service at Location B> Safer than before Safer than before (35.1%) (43.7%) No diference No diference (56.3%) (64.9%) <Is the service necessary at Location A?> <Is the service necessary at Location B?> 91 78 Yes (92.9%) Yes (91.8%) No No 7 (7.1%) (8.2%) Necessary Necessary Not necessary Not necessary <Is it safer than before at Location C?> (85.1%) <Awareness of service at Location C> (14.9%) Safer than before 77 74 No diference (51.0%) (49.0%) <Is the service necessary at Location C?> Yes (94.6%) No (5.4%) Necessary Not necessary Figure 12: Perception of service at each location. following formula: Degree of freedom � (the number of We obtained two χ values of 1.6444 and 0.4911, with as- rows − 1) (the number of columns − 1). We established the sociated p values of 0.4395 and 0.7823, respectively. Since two diferent null hypotheses that the proportion among the these p values are greater than the signifcance level of 0.05, three groups is the same for the following questions: (i) “Do we do not reject the two hypotheses. Te report states that you think that the service has increased safety at this lo- there is no statistically signifcant diference among the cation?” and (ii) “Do you think that the service is necessary locations in terms of service perception. Each location at this location?.” If the calculated χ -value is greater than the represents scenarios that can happen on residential roads. critical value from the χ -distribution, we must reject the Accordingly, it is demonstrated that the proposed service null hypothesis. Tis implies that at least one proportion can be implemented in many locations on residential roads difers considerably from another proportion among groups. and have the same efect regardless of location from a user’s 14 Journal of Advanced Transportation Table 5: Chi-square test results. Number of responses (%) Results Question Location Yes No Total χ df p value A 77 (78.6%) 21 (21.4%) 98 (1) Do you think that the service has increased safety at this location? B 66 (77.6%) 19 (22.4%) 85 1.6444 2 0.4395 C 63 (85.1%) 11 (14.9%) 74 A 91 (92.9%) 7 (7.1%) 98 (2) Do you think that the service is necessary at this location? B 78 (91.8%) 7 (8.2%) 85 0.4911 2 0.7823 C 70 (94.6%) 4 (5.4%) 74 Table 6: Suggestions for improving trafc safety at each location, based on open-ended questions. Ideas for Ideas for Ideas for % % % Location A Location B Location C Use acoustic speakers 20 Need advertisements 19 Use acoustic speakers 25 Need advertisements 16 Use acoustic speakers 11 Need trafc safety signs 22 Need convex mirrors 13 Prohibit illegal roadside parking 11 Install automatic roadblock 14 perspective. Te results of the chi-square test are summa- improving trafc safety on KAIST Seoul Campus. In ad- rized in Table 5. dition, they stated that the campus was safer than before the Finally, we asked respondents for suggestions on im- service installation, with 78.6%, 77.6%, and 85.1% proving trafc safety at each location. Te suggestions are responding positively for Locations A, B, and C, respectively. summarized in Table 6. One of the most recommended Tese results, combined with the high satisfaction reported approaches at almost all locations was to use an acoustic by survey respondents, suggest that our service can be ap- speaker to deliver warning information. Some respondents plied to various areas that are typically considered residential suggested that it would be more efective if both a visual and roads. Te service implementation is expected to improve an audible warning were utilized concurrently. Another trafc safety and reduce fatalities that arise due to blind spots. suggestion was to promote the service to campus members. Indeed, in Locations A, B, and C, 64.9%, 56.3%, and 49.0% of Te generalizability of these results is subject to certain respondents, respectively, knew that the service was in limitations. For instance, evaluating the proposed service operation. Except for the use of acoustic speakers, Location was conducted based only on a survey. A natural progression C’s suggestions difered from those of the other locations. of this work is to evaluate the service using before-and-after Te characteristics of Location C included trafc safety signs surrogate data, such as the frequency of two- and three- for one-way driving and an automatic roadblock that pre- second TTC events. We are currently collecting the relevant vents vehicles from traveling in the incorrect direction. data for this purpose. In addition, TTC may also vary based on given infrastructure, nature, and human environments. In this study, a constant TTC threshold was adopted with 4.Conclusions a safety margin, which could yield unnecessary false posi- tives in some situations. Terefore, we should conduct In this study, we propose a collision risk warning service for a sensitivity analysis for TTC with respect to diferent residential roads based on risk assessment. In contrast to location-specifc environments to determine the optimal earlier research, this service combines CCTVs and radar to TTC for each location. Moreover, using acoustic speakers to detect items precisely and quickly. We use an elliptical alert vehicles and pedestrians appeared to be the most trajectory prediction approach to predict unknown pedes- suggested approach for enhancing trafc safety at all loca- trian behaviors. Te major and minor axes of the ellipse were tions. According to the respondents, they believed that derived using CCTV data on actual pedestrian trajectories. combining visual and audible warnings would provide Furthermore, we use TTC to identify collision risks in a more efective warning to those using a smartphone and/or vehicle-vehicle and vehicle-pedestrian cases. An LED sign earphones. Te results are consistent with [43], which board is used to provide risk warnings to vehicles and pe- supported the idea that multimodal warning services have destrians. Te proposed service was provided in three sit- potential advantages in various situations, such as when uations: illegal roadside parking, unprotected left turn, and people are using smartphones or are engaged in distracted wrong-way driving on residential roads. driving. We applied our service to three locations on KAIST Seoul Campus in Korea. To evaluate service efects, we conducted a survey and analyzed the safety efects from the Data Availability user’s perspective. Using a set of questions, we investigated respondents’ satisfaction with the service. 76.9% of re- Te data used to support the fndings of this study are spondents reported that the service contributed to available from the corresponding author upon request. Journal of Advanced Transportation 15 [13] B. Noh and H. Yeo, “A novel method of predictive collision Conflicts of Interest risk area estimation for proactive pedestrian accident pre- vention system in urban surveillance infrastructure,” Trans- Te authors declare that they have no conficts of interest. portation Research Part C: Emerging Technologies, vol. 137, Article ID 103570, 2022. Authors’ Contributions [14] W. Yang, X. Zhang, Q. Lei, and X. Cheng, “Research on longitudinal active collision avoidance of autonomous Aya Selmoune and Jeongin Yun contributed equally to emergency braking pedestrian system (AEB-P),” Sensors, this paper. vol. 19, no. 21, 2019. [15] R. Matsumi, P. Raksincharoensak, and M. Nagai, “Autono- Acknowledgments mous braking control system for pedestrian collision avoid- ance by using potential feld,” IFAC Proceedings Volumes, Tis work was supported by the Ministry of the Interior and vol. 46, no. 21, pp. 328–334, 2013. Safety (MOIS), Republic of Korea (grant nos. 2021-MOIS41- [16] A. Høye and A. Laureshyn, “SeeMe at the crosswalk: before- 001-00000000-2022 (development of trafc safety risk after study of a pedestrian crosswalk warning system,” Transportation Research Part F: Trafc Psychology and Be- warning technology for blind spots in living roads)). haviour, vol. 60, pp. 723–733, 2019. [17] D. Ka, D. Lee, and H. Yeo, “Development of predictive pe- References destrian collision warning service considering pedestrian characteristics,” Te Journal of Te Korea Institute of In- [1] WHO, “WHO road trafc injuries,” 2022, https://www.who. telligent Transport Systems, vol. 18, no. 3, pp. 68–83, 2019. int/news-room/fact-sheets/detail/road-trafc-injuries. [18] J. J. Anaya, P. Merdrignac, O. Shagdar, F. Nashashibi, and [2] PWC, “A guide for policy makers: on reducing road fatalities,” J. E. Naranjo, “Vehicle to pedestrian communications for 2022, https://www.pwc.com/m1/en/publications/guide-on- protection of vulnerable road users,” in Proceedings of the reducing-road-fatalities.html. IEEE Intelligent Vehicles Symposium Proceedings, Dearborn, [3] “OECD International Transport Forum: Road injury acci- MI, USA, June 2014. dents,” 2022, https://stats.oecd.org/Index.aspx?DataSetCode= [19] K. Dhondge, S. Song, B. Choi, and H. Park, “WiFiHonk: ITF_ROAD_ACCIDENTS. smartphone-based beacon stufed WiFi Car2X- [4] M. W. Adler and R. Ahrend, “Trafc safety in South Korea: communication system for vulnerable road user safety,” in understanding the vulnerability of elderly pedestrians,” OECD Proceedings of the IEEE 79th Vehicular Technology Conference Regional Development Working Papers, Trafc Safety in Korea, (VTC Spring), pp. 1–5, Seoul, Korea, May 2014. OECD Publishing, Seoul, Korea, 2017. [20] T. Wang, G. Cardone, A. Corradi, L. Torresani, and [5] “OECD International Transport Forum: Road safety annual A. T. Campbell, “Walksafe: a pedestrian safety app for mobile report 2021 korea,” 2022, https://www.itf-oecd.org/sites/ phone users who walk and talk while crossing roads,” in default/fles/korea-road-safety.pdf. Proceedings of the Twelfth Workshop on Mobile Computing [6] J. Lim, S. Lee, J. Choi, and S. Joo, “Te comparative study on Systems & Applications, pp. 1–6, San Diego, CA, USA, Feb- travel behavior and trafc accident characteristics on a com- ruary 2012. munity road-with focus on Seoul metropolitan city,” Journal [21] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only of the Korean Society of Civil Engineers, vol. 36, no. 1, Look once: unifed, real-time object detection,” in Proceedings pp. 97–104, 2016. of the IEEE Conference on Computer Vision and Pattern [7] S. Tak, S. Kim, D. Lee, and H. Yeo, “A comparison analysis of Recognition, pp. 779–788, las vegas, NV, USA, June 2016. surrogate safety measures with car-following perspectives for [22] F. Zhou, H. Zhao, and Z. Nie, “Safety helmet detection based advanced driver assistance system,” Journal of Advanced on YOLOv5,” in Proceedings of the IEEE International Con- Transportation, vol. 2018, Article ID 8040815, 14 pages, 2018. ference on Power Electronics, Computer Applications [8] S. o. Son, J. Park, C. Oh, and C. Yeom, “An algorithm for (ICPECA), pp. 6–11, Shenyang, China, January 2021. detecting collision risk between trucks and pedestrians in the [23] R. Padilla, S. L. Netto, and E. A. Da Silva, “A survey on connected environment,” Journal of Advanced Trans- performance metrics for object-detection algorithms,” in portation, vol. 2021, Article ID 9907698, 9 pages, 2021. Proceedings of the 2020 international conference on systems, [9] L. Zhang, K. Yuan, H. Chu et al., “Pedestrian collision risk signals and image processing (IWSSIP), pp. 237–242, IEEE, assessment based on state estimation and motion prediction,” Niteroi, Brazil, July 2020. IEEE Transactions on Vehicular Technology, vol. 71, no. 1, [24] R. Solovyev, W. Wang, and T. Gabruseva, “Weighted boxes pp. 98–111, 2022. fusion: ensembling boxes from diferent object detection [10] Y. Wu, M. Abdel-Aty, O. Zheng, Q. Cai, and L. Yue, “De- models,” Image and Vision Computing, vol. 107, Article ID veloping A crash warning system for the bike lane area at 104117, 2021. intersections with connected vehicle technology,” Trans- [25] I. Culjak, D. Abram, T. Pribanic, H. Dzapo, and M. Cifrek, “A portation Research Record, vol. 2673, no. 4, p. 47, 2019. [11] S. Zhang, M. Abdel-Aty, Y. Wu, and O. Zheng, “Modeling brief introduction to OpenCV,” in Proceedings of the 35th International convention MIPRO, pp. 1725–1730, Opatija, pedestrians’ near-accident events at signalized intersections using gated recurrent unit (GRU),” Accident Analysis & Croatia, May 2012. [26] B. Noh, W. No, and D. Lee, “Vision-based overhead front Prevention, vol. 148, Article ID 105844, 2020. [12] S. Zhang, M. Abdel-Aty, Y. Wu, O. Zheng, and J. Ugan, point recognition of vehicles for trafc safety analysis,” in Proceedings of the 2018 ACM International Joint Conference “Prediction of pedestrian-vehicle conficts at signalized in- tersections based on long short-term memory neural net- and 2018 International Symposium on Pervasive and Ubiq- uitous Computing and Wearable Computers, pp. 1096–1102, work,” Accident Analysis & Prevention, vol. 148, Article ID 105799, 2020. Singapore, October 2018. 16 Journal of Advanced Transportation [27] S. S. Pathan, A. Al-Hamadi, and B. Michaelis, “Intelligent [42] T. M. Franke, T. Ho, and C. A. Christie, “Te chi-square test: feature-guided multi-object tracking using kalman flter,” in often used and more often misinterpreted,” American Journal of Evaluation, vol. 33, no. 3, pp. 448–458, 2012. Proceedings of the 2nd International Conference on Computer, [43] F. Biondi, D. L. Strayer, R. Rossi, M. Gastaldi, and C. Mulatti, Control and Communication, pp. 1–6, Karachi, Pakistan, “Advanced driver assistance systems: using multimodal re- February 2009. dundant warnings to enhance road safety,” Applied Ergo- [28] X. Li, K. Wang, W. Wang, and Y. Li, “A multiple object nomics, vol. 58, pp. 238–244, 2017. tracking method using kalman flter,” in Proceedings of the IEEE International Conference on Information and Automa- tion, pp. 1862–1866, Harbin, China, June 2010. [29] N. S. Punn, S. K. Sonbhadra, S. Agarwal, and G. Rai, “Monitoring COVID-19 social distancing with person de- tection and tracking via fne-tuned YOLO v3 and Deepsort techniques,” 2020, https://arxiv.org/abs/2005.01385. [30] A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in Proceedings of the 2016 IEEE international conference on image processing (ICIP) 2016, pp. 3464–3468, Phoenix, AZ, USA, September 2012. [31] G. Shu, A. Dehghan, O. Oreifej, E. Hand, and M. Shah, “Part- based multiple-person tracking with partial occlusion han- dling,” in Proceedings of the 2012 IEEE Conference on Com- puter Vision and Pattern Recognition 2012, pp. 1815–1821, Rhode Island, RI, USA, June 2012. [32] K. Bernardin and R. Stiefelhagen, “Evaluating multiple object tracking performance: the clear mot metrics,” EURASIP Journal on Image and Video Processing, vol. 2008, Article ID 246309, 10 pages, 2008. [33] K. S. Oh, S. Y. Park, J. H. Seo, G. H. Lee, and K. S. Yi, “Laser- Scanner-based stochastic and predictive working-risk- assessment algorithm for excavators,” Journal of Drive and Control, vol. 13, no. 4, pp. 14–22, 2016. [34] M. Seo, H. Shin, H. Jeong, and J. Chae, “Development of an object collision detection algorithm for prevention of collision accidents on living roads,” Journal of Drive and Control, vol. 19, no. 3, pp. 23–31, 2022. [35] R. Miller and Q. Huang, “An adaptive peer-to-peer collision warning system Vehicular Technology Conference,”vol. 1, pp. 317–321, in Proceedings of the IEEE 55th Vehicular Technology Conference. VTC Spring, vol. 1, pp. 317–321, IEEE, Birmingham, AL, USA, May 2020. [36] F. Jimenez, ´ J. E. Naranjo, and F. Garc´ıa, “An improved method to calculate the time-to-collision of two vehicles,” International Journal of Intelligent Transportation Systems Research, vol. 11, no. 1, pp. 34–42, 2013. [37] R. Van Der Horst and J. Hogema, Time-to-collision and Collision Avoidance Systems, in Proceedings of the 6th ICTCT Workshop Salzburgna, Salzburg, Te Netherlands, 1993. [38] S. Das and A. K. Maurya, “Defning time-to-collision thresholds by the type of lead vehicle in non-lane-based trafc environments,” IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 12, pp. 4972–4982, 2020. [39] F. Bella and R. Russo, “A collision warning system for rear- end collision: a driving simulator study,” Procedia-social and behavioral sciences, vol. 20, pp. 676–686, 2011. [40] U. Durrani, C. Lee, and D. Shah, “Predicting driver reaction time and deceleration: comparison of perception-reaction thresholds and evidence accumulation framework,” Acci- dent Analysis & Prevention, vol. 149, 2021. [41] M. L. McHugh, “Te chi-square test of independence,” Bio- chemia Medica, vol. 23, no. 2, pp. 143–149, 2013. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Advanced Transportation Hindawi Publishing Corporation

Development of a Residential Road Collision Warning Service Based on Risk Assessment

Download PDF
16 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/development-of-a-residential-road-collision-warning-service-based-on-XdXgaAwRl2

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
Hindawi Publishing Corporation
ISSN
0197-6729
eISSN
2042-3195
DOI
10.1155/2023/7496377
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Advanced Transportation Volume 2023, Article ID 7496377, 16 pages https://doi.org/10.1155/2023/7496377 Research Article Development of a Residential Road Collision Warning Service Based on Risk Assessment 1 2 3 4 Aya Selmoune , Jeongin Yun , Myoungkook Seo , Hyeokhyeon Kwon , 5 2 Changhee Lee , and Jinwoo Lee Jiangsu Key Laboratory of Urban ITS, Southeast University, Nanjing 210000, China Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon 34051, Republic of Korea Smart Engineering Laboratory, Korea Construction Equipment Technology Institute, Gunsan 54004, Republic of Korea Research and Development Center, Pintel Incorporated, Seoul 06729, Republic of Korea Transportation Policy Division, Daejeon Metropolitan City Hall, Daejeon 35242, Republic of Korea Correspondence should be addressed to Jinwoo Lee; lee.jinwoo@kaist.ac.kr Received 9 August 2022; Revised 4 February 2023; Accepted 11 February 2023; Published 16 March 2023 Academic Editor: Jose E. Naranjo Copyright © 2023 Aya Selmoune et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Pedestrians are more likely to be seriously injured in vehicle collisions. In fact, multiple collisions between vehicles and pedestrians occur on residential roads that lack street-to-sidewalk dividers and have numerous blind spots. Traditional trafc safety features and equipment, such as speed bumps and trafc signs, are not always sufcient to prevent pedestrian accidents on such residential roads. Terefore, we suggest a collision risk warning service for residential roads as a solution to this issue. We use CCTVs with computer vision techniques and radar to accurately detect objects in real-time and to trace their trajectories. In addition, we employ a time-to-collision-based method to identify dangerous situations. Te service warns drivers and pedestrians about hazardous situations using a light-emitting diode sign board. We applied our service to three diferent roads on a university campus in Seoul, Korea, and then conducted a user survey to evaluate the service. In summary, more than 90% of respondents stated that the service was necessary for these specifc locations, and 76.9% noted that the service signifcantly contributed to trafc safety on the campus. Tis implies that the proposed service improved trafc safety and can be applied to various locations on residential roads. [6]. Pedestrians are the most likely to be seriously injured in 1.Introduction vehicle collisions. Traditional trafc safety features and Approximately 1.3 million people die annually because of equipment, such as speed bumps and trafc signs, are not trafc accidents [1]. Some governments and agencies in always sufcient for preventing pedestrian accidents in blind many countries have tried to reduce trafc accidents by spots on residential roads. Particularly when pedestrians abruptly exit from parked vehicles on roads, drivers are implementing safety education and policies such as pro- moting trafc rules and enforcing speed limits [2]. As a result unable to respond appropriately, and trafc accidents are of these eforts, trafc fatalities in most developed countries highly possible. in the OECD have decreased substantially. For example, Several technologies have been developed to prevent Korea reduced road fatalities by 26.4% from 2017 to 2020 [3]. vehicle-pedestrian collisions. Tey are based on algorithms However, safety issues of pedestrians remain a concern [4]. that identify objects, predict their trajectories, and determine Pedestrian fatalities in Korea accounted for 35% of total whether or not a collision risk exists. Te algorithms can be fatalities [5]. More than half of pedestrian fatalities occur on divided into two categories depending on how the collision residential roads without separation of streets and sidewalks risk is determined. First, some algorithms employ surrogate 2 Journal of Advanced Transportation drivers. Few services considered a pedestrian perspective. safety measures (SSMs) to recognize the presence of po- tentially dangerous situations based on whether predicted One study developed a system that recognized dangerous situations and provided information to pedestrians via their trajectories of objects overlap. Using microscopic trafc characteristics such as vehicle speed, acceleration, time smartphones [20]. However, it was inaccurate and in- headway, and space headway, an SSM method assesses the efective in that object detection was conducted only by collision risk of particular trafc scenarios [7]. SSMs, such as cameras on smartphones. In addition, few studies evaluated time-to-collision (TTC) and post encroachment time (PET), the efects of proposed algorithms in the feld. Most algo- have been widely used to evaluate trafc safety performance rithms were evaluated based on simulations or feld pro- and identify potential accident risks [8–13]. One study as- totype tests, and accuracy was only verifed through a confusion matrix. sumed a connected environment in which pedestrians and vehicles shared real-time location information using IoT In the present study, we propose a safety service framework that provides risk information to both vehicles devices. According to object locations, velocity, relative distance, angle, and TTC, dangerous situations were de- and pedestrians. Te proposed framework utilizes RSE such as CCTVs and radar to detect objects using a deep learning termined [8]. In another study, an algorithm was developed using onboard cameras in vehicles. Potential collision areas method. Ten, the algorithm uses SSMs to identify whether were defned by the minimum TTC from the predicted the current situation is dangerous. If the situation is unsafe, movements of ego vehicles and pedestrians [9]. In addition, a light-emitting diode (LED) sign board gives warning in- in a connected vehicle environment, a crash warning system formation to both vehicles and pedestrians to avoid a po- was developed for bike lane areas. PET was used to identify tential collision. Tus, the service alerts drivers and potential areas of interaction between vehicles and bicycles pedestrians at the same time. To evaluate the safety efects of the proposed service, we implemented and operated it on- [10]. Most algorithms were verifed as simulation-based or autonomous platforms. Te second set of algorithms pre- site. We conducted a survey to investigate user satisfaction. Te remainder of this paper is structured as follows: the dicts risk situations using deep learning methods [11–13]. After an algorithm is trained using prior data labeled by service description section presents the overall framework. Te application and evaluation section introduces the study SSMs as risk situations, it predicts whether a given situation is dangerous. Te gated recurrent unit method was used to site and presents the evaluation. In the last section, we predict collision risk at a signalized intersection [11]. Sim- summarize this study and discuss possible future research ilarly, long-short term memory (LSTM) was used to predict directions. risk situations [12]. In some cases, deep learning methods were used for trajectory estimation to predict risk situations. 2. Service Description One study proposed a collision risk area estimation system at unsignalized crosswalks. Te system used LSTM to predict We propose a collision risk warning service procedure, as object trajectories and then conducted statistical inferencing depicted in Figure 1. Tis service is a proactive counter- to predict collision risk areas [13]. measure against vehicle-vehicle or vehicle-pedestrian col- As soon as a potentially hazardous situation is identifed, lisions. Tere are four steps: Step 1 object detection through various warning services are provided. Tis warning in- CCTV and radar; Step 2 trajectory prediction of detected formation can be divided into three categories. First, in- objects; Step 3 collision risk identifcation based on predicted formation is provided by vehicles. Augmented reality (AR) trajectories; and Step 4 collision risk warning, if any. Here, on the heads-up display in vehicles was employed to display the current time is t , and the previous point one time step warning information. In addition to AR, an audio warning before and the future point n time steps after are denoted by was immediately provided [10]. Several active pedestrian t and t . −1 n collision avoidance systems did not give alerts but instead automatically controlled the vehicles [14, 15]. Te second method is to provide information to vehicles from roadside 2.1. Object Detection. We use CCTV and radar equipment to equipment (RSE). For example, amber fashing lights were detect vehicles and pedestrians. One of the detecting algo- activated when pedestrians were approaching or crossing rithms is you only look once (YOLO) [21], which has been crosswalks so that the drivers could perceive them [16]. Te used in various felds for real-time detection. We employ third method is to use infrastructure-to-vehicle (I2V), ve- a YOLO v5-based algorithm. YOLO v5 is faster and more hicle-to-pedestrian (V2P), and vehicle-to-everything (V2X) accurate than its previous versions [22]. To account for the communication. One study utilized I2V communication to characteristics of residential roads, we need residential-road- give warning information to vehicles from RSE [17]. Several specifc training datasets, which are distinguished from studies developed V2P and V2X communication-based general road datasets. Terefore, we used 150 hours of video warning services in Wi-Fi environments [18, 19]. How- data from CCTV cameras installed on residential roads in ever, in the current state, the communication-based safety Guro-gu, Seoul. We trained for various environments such warning method has problems regarding latency and as lighting and weather conditions as well as situations stability. involving numerous objects such as pedestrians carrying Most systems were developed from the perspective of umbrellas, as shown in Figure 2. With the trained model, vehicles. Based on cameras or radar sensors in vehicles and objects can be accurately identifed in real-time as pedes- CCTVs in RSE, warning information was provided to trians, motorcycles, bicycles, vehicles, and personal mobility Journal of Advanced Transportation 3 Step 1. Object detection Transmitted pulse Reactive pulse Radar CCTV Step 2. Trajectory prediction t t t t t t –1 0 n –1 0 n Step 3. Collision risk identifcation Dangerous Not dangerous Step 4. Collision risk warning Figure 1: Overall structure of collision risk warning service. devices, even under severe lighting and weather conditions. in Table 1. At the 50% level of Intersection over Union, In Figure 3, the training results are shown with an example defned as the degree of overlap between ground truth and site at two diferent time points compared to the identif- prediction regions [23, 24], the detection rate for pedes- cation without the training. Te overall accuracy is presented trians, motorcycles, bicycles, vehicles, and personal mobility 4 Journal of Advanced Transportation Table 1: Post-training object detection accuracy. Pedestrian Motorcycle Bicycle Vehicle Personal mobility device Proposed algorithm 2573/2590 2213/2216 2994/2995 341/343 934/942 (Detection rate) (99.34%) (99.86%) (99.97%) (99.42%) (99.15%) Number of accurately detected objects/total number of objects. Journal of Advanced Transportation 5 devices was higher than 99%. In addition, radar is employed Multiple object tracking accuracy (MOTA) is used to to complement CCTVs. Tey provide precise locations and evaluate the accuracy of object-tracking algorithms [32]. MOTA is the most prevalent indicator used to measure speeds, which are difcult to collect with CCTVs. With these two complementary devices, accurate and precise real-time a tracker’s performance. Its value may be determined using object detection is achieved. equation (3). IDSW + FN + FP 2.2. Trajectory Prediction. Future trajectories of objects are (3) MOTA � 1 − , GT predicted based on their previous coordinates that we can track. In this study, perspective transformation and Kalman flter are where ground truth (GT) is the total number of ground truth used for tracking objects. Coordinates of objects detected by objects, identity switching (IDSW) represents the number of CCTVs are transformed into overhead perspectives to measure ID switches in the video stream, false negative (FN) indicates exact locations. We employ the perspective matrix in Open CV a missed detection, and false positive (FP) means an in- to convert the coordinates from the videos to overhead co- accurate detection. Based on MOTA, we observe that the ordinates [25, 26]. Kalman flter involves repeating the pre- proposed algorithm results in higher accuracy than Deep- diction step and correction step of trajectories [27]. In the SORT. Mostly tracked targets (MT) and mostly lost targets prediction step, the next position of the object in the current time (ML) are the number of tracked and lost objects, re- is estimated based on the information collected about the object spectively. Te proposed algorithm has higher and lower already being tracked, as in equation (1). values of MT and ML, respectively, than DeepSORT, which x � Ax + w , (1) are desirable. Furthermore, the proposed object-tracking k k−1 k−1 algorithm has a higher FPS than DeepSORT since we do where x is the state vector representing the object’s dynamic k not use the computationally burdening Hungarian matching behavior at a discrete time index k; A is the transition matrix algorithm. In summary, the proposed tracking algorithm at time index k − 1 to k; and vector w is the noise fol- k−1 outperforms DeepSORT. lowing normal probability distribution N(0, Q) with zero Ten, based on the tracking data, we predict vehicle and mean and covariance matrix Q. In the correction step, the pedestrian trajectories. First, we classify straight and curved previously-predicted position is compared to the position trajectories based on whether the angle of the previous measured by CCTVs. To modify the object position, a weight trajectories is smaller than the angle that we predetermine, called Kalman gain K is used, which indicates the ratio of k θ > 0 (unit: radians). We estimate a vehicle’s tendency to set the error of the predicted object position to the error of the move based on the angle diference between previous points. object position measured by the object detection algorithm. Ten, we refect this tendency in the trajectory prediction. If K has a range from zero to one, and it is infuenced by more we set the time index k to zero for the current time, the accurate values between the predicted position and the ∗ current position is P at x , and the previous positions at 0 k�0 measured position, as stated in equation (2). two and one time steps before are P and P , respectively, −2 −1 − − ∗ ∗ x 􏽢 � x 􏽢 + K z − Hx 􏽢 􏼁 , (2) i.e., the coordinates of P and P are x and x . Te k k k k k −2 −1 k�−2 k�−1 �������→ �����→ known angle between P P and P P is θ (unit: ra- − −2 −1 −1 0 −1 􏽢 􏽢 where x is a posteriori estimated state; x is a priori es- k k dians). Te tendency angle θ (unit: radians) is calculated in timate; z is the observed measurement; and H is the equation (4). measurement matrix at time k − 1  to  k. After the repeated execution, we update K and fnd the optimal state (x ) that k 0 minimizes the error between the estimated state and the θ � 􏽘 θ ω , (4) 0 k k measured state [28]. past k�−T Te object-tracking algorithm proposed in this where ω is a weight factor that considers the angle error, s.t., study was compared to DeepSORT, which is a deep past ω � 1, and T is the time window length of the past learning-based method for tracking objects [29, 30]. k�−T k data, i.e., we consider the tracking data at DeepSORT consists of four key components: detection, past past k ∈ −T , −T + 1, . . . , −1, 0 . 􏼈 􏼉 estimation, data association, and generation and de- We consider a short prediction period between k � 0 and letion of tracking objects. In DeepSORT, Kalman flter future future past k � T , where T ≪ T . Tus, it can be reasonable is used in the estimation stage, and a Hungarian to assume that a vehicle with |θ | < θ keeps moving straight 0 set matching algorithm is employed in the data associa- future future during T . At future time point n ∈ 􏼈1, . . . , T − 1􏼉, tion stage [30]. Te major diference between Deep- the center location of the straight moving vehicle, denoted SORT and the proposed algorithm is twofold. First, we �����→ P , is on the straight line extended from P P and the n −1 0 advanced the Kalman flter algorithm. Second, due to ����→ the Hungarian algorithm’s prohibitive computational 􏽢 􏽢 distance from P to P , |P P |, is the multiplication of the 0 n 0 n cost, we developed an original matching algorithm average vehicle speed and the time diference between t and instead of using the Hungarian algorithm. For com- t . If |θ | ≥ θ , the center of a vehicle predicted at time point n 0 set parison, we used the Oxford Town Centre dataset, 􏽢 􏽢 n, P , is found based on P , θ , and the average vehicle n n−1 n−1 which is commonly employed to assess object-tracking speed. Particularly, P is calculated using P and θ . Similar 1 0 0 performance [31]. Te comparison results are shown in n −1 􏽢 􏽢 to equation (4), θ is defned as 􏽐 θ ω +􏽐 θ ω Table 2. n k�0 k k−n k�n−T k k−n 6 Journal of Advanced Transportation (a) (b) (c) (d) (e) (f) Figure 2: Various environments for training model. Te model was trained in various lighting conditions, including clear (a, b, c), snowy (d), and rainy weather (e, f). future range that the pedestrians reach varies. Te estimated for all n ∈ 􏼈1, . . . , T 􏼉. Te spatial range of a vehicle, major and minor axes of the ellipse are determined using predicted at n, is defned to have its center at P , and its equation (5), where the parameters were tuned based on boundary is determined based on the actual vehicle size. the actual data. Compared to vehicles, pedestrians have relatively uncertain and inconsistent movement characteristics. major ⎧ ⎪ 1.6 d − 0.89, 􏼐e ≥ 1.0􏼑, ⎪ n n Tus, we use an elliptical trajectory prediction approach major e � to account for stochastic pedestrian trajectories, and we ⎪ major 1.0, 􏼐e < 1.0􏼑, consider the ellipse as the future spatial range of a pe- destrian’s location [33]. We estimate the moving di- rection of pedestrians based on the previous directions in ⎧ ⎪ 2.0, d ≥ 3.0 􏼁 , (5) the same method for vehicles, as described in equation minor e � (4). We estimate the major and minor axes of an ellipse ⎪ 1 using actual pedestrian path data collected from CCTVs ⎪ ⎩ d + 1.0, d < 3.0 , n n on residential roads [34]. Estimation results depend on the time point t for all at which we predict from the k�n 􏼌 􏼌 􏼌 􏼌 future 􏽢 􏼌 􏼌 present time, n > 0. We can fnd P of the center of an 􏼌 􏼌 d � P P , ∀n ∈ 1, . . . , T , 􏽮 􏽯 n 􏼌 􏼌 n 0 n ellipse for a pedestrian in a similar way to fnding that for major a vehicle with |θ | ≥ θ . Depending on the moving dis- where e (unit: meters) is the major axis of the ellipse at 0 set minor tance d (unit: meters) from the current time point t to future time point n, and e (unit: meters) is the minor n 0 n the future point t , defned as |P P |, the possible spatial axis of the ellipse at future time point n. n 0 n Journal of Advanced Transportation 7 (a) (b) Figure 3: Detection performance improvement: (a) before and (b) after additional training. After additional training, the model detects a pedestrian carrying umbrella and distinguishes a personal mobility device from a pedestrian. Table 2: Object-tracking performance comparison. MOTA IDSW FN FP MT ML GT FPS (%) DeepSORT 79.90 1024 5667 2753 114 2 46985 68.65 Proposed tracking 81.70 743 3590 4258 136 1 46945 333.79 algorithm Te predicted trajectories are graphically illustrated in coordinate of P . Te MAE was calculated to be between 0.09 Figure 4. and 0.65 meters. Te accuracy is lowered as the curvature We tested the accuracy of the trajectory prediction model and speed increase. using example trajectories of three pedestrians and two vehicles, as depicted in Figure 5. Te length of each time period is one 2.3. Collision Risk Identifcation. Once future trajectories of frame, and for each prediction, we plotted the center of a pe- objects intersect, we get one intersecting point destrian’s ellipse or a vehicle at n � 3 (after three seconds from (q 􏽢 􏽢 , s ) of two objects a and b, and two collision time each prediction time point). Te prediction trial is indexed by i intersect intersect (CT) to reach the intersecting point from each object at the and the total number of trials is I. We conducted 857, 396, 899, current location. q 􏽢 and 􏽢 s are calculated by equation 330, and 324 trials for pedestrian #1, pedestrian #2, pedestrian intersect intersect (7) [35]. #3, vehicle #1, and vehicle #2, respectively. Te test results are b a presented in Figure 6 and Table 3. Te unit of the graphs in b a b a 􏽢 􏽢 􏼐s − s 􏼑 − 􏼒q tan θ − q tan θ 􏼓 0 0 0 0 0 0 Figure 6 is in pixels, and the horizontal and vertical lengths of q 􏽢 � , intersect a b a pixel are 0.09 meters and 0.11 meters, respectively. For the 􏽢 􏽢 tan θ − tan θ 0 0 accuracy measure, we use mean absolute error (MAE) calculated (7) by equation (6). b a b a b a 􏽢 􏽢 􏼐q − q 􏼑 − 􏼒s cot θ − s cot θ 􏼓 􏼌 􏼌 0 0 0 0 0 0 􏼌 􏼌 I 􏼌 i i 􏼌 􏼌 􏼌 x 􏽢 − x 􏽢 s � , 􏼌 􏼌 i�1 3 3 intersect a b (6) 􏽢 􏽢 MAE � , cot θ − cot θ 0 0 i i where q 􏽢 and 􏽢 s represent the longitudinal and where x 􏽢 and x are the predicted and the actual locations of intersect intersect 3 3 latitudinal coordinates of the intersecting point, respectively; the object three seconds later than the current time point of a a b b th i (q , s ) and (q , s ) represent the current coordinates of the i prediction trial, respectively. Specifcally, x 􏽢 is the 0 0 0 0 3 8 Journal of Advanced Transportation |θ | < θ 0 set –1 P ⌃ –1 P θ n –1 –2 Straight line –1 1 Vehicle’s spa spat ti ial ra al ran ng ge e |θ | ≥ θ 0 set θ P –1 P –1 –1 –2 Curved line P edestrian’s spatial range (a) (b) Figure 4: Trajectory prediction method: (a) trajectory classifcation—straight and curved trajectories are classifed based on whether the angle of the previous trajectories is smaller than the angle predetermined. (b) Trajectory predictions—vehicle and pedestrian trajectories are predicted as rectangles and ellipses, considering the actual vehicle size and stochastic pedestrian trajectories, respectively. Pedestrian trajectory Vehicle trajectory Figure 5: Test scenario description. Trajectories of three pedestrians and two vehicles were tested. a b a 􏽢 􏽢 􏽢 objects a and b, respectively; and tan θ , tan θ , cot θ , and considering the diference between the center point and the 0 0 0 cot θ represent the current tangent and cotangent values of spatial range boundary of each object. objects a and b. With q 􏽢 , 􏽢 s , and the objects’ speeds intersect intersect If the spatial ranges of two objects successively and directions, we can calculate CTs for both objects, overlap at more than three intervals, the time interval a b 􏽣 􏽣 CT   and  CT [36]. We compare the two CTs, and the when the two spatial ranges frst overlap is defned as the a b predicted TTC. Tis indicator assumes that the involved 􏽣 􏽣 smaller one, CT ( ≡ min CT , CT , unit: seconds), is 􏼚 􏼛 min objects do not recognize the risk and there is no urgent used in determining whether the spatial ranges overlap in maneuver to avoid it in a following short period of time. increments of 0.25 seconds starting from one second earlier We compare the predicted TTC with a TTC threshold to than CT (i.e., CT − 1 + 0.25δ  for  δ � 0, 1, 2, . . .) min min identify whether a collision risk exists [37–39]. If it is Journal of Advanced Transportation 9 (unit: pixels) Pedestrian #1 Pedestrian #2 Pedestrian #3 620 640 660 680 700 0 500 1000 1500 620 640 660 680 700 0 300 Expected route Expected route Expected route Real route Real route Real route Vehicle #1 Vehicle #2 620 640 660 680 700 0 200 400 600 800 1000 0 370 900 420 Expected route Expected route Real route Real route Figure 6: Trajectory prediction test results. Table 3: Trajectory prediction accuracy. Pedestrian Vehicle #1 #2 #3 #1 #2 MAE 0.09 0.39 0.18 0.54 0.65 (unit: meters) Current state Future state: After CT – 1 + 0.25δ seconds (δ = 0,1,2,...) min CT seconds ... CT seconds Object a ... CT – 1.00 CT – 0.75 CT – 0.50 CT – 0.25 CT Object b min min min min min a b here, predicted TTC = CT – 0.50 (units: seconds) min CT ≡ min {CT , CT } min Spatial range overlap (a) (b) a b 􏽣 􏽣 Figure 7: Determination of risk existence and prediction of time-to-collision (TTC): (a) CT is determined based on CT and CT . (b) If min two spatial ranges overlap successively more than three intervals, the predicted TTC is calculated as the frst overlap, and if the predicted TTC is smaller than the TTC threshold, the collision risk is regarded to exist. 10 Journal of Advanced Transportation smaller than the TTC threshold, the collision risk is status and the service on the site, as shown in Figure 10. regarded to exist. Te entire process of the TTC calcu- 32.5% of respondents reported that they had encountered unsafe situations on KAIST Seoul Campus, with 44.9% in A, lation is depicted in Figure 7. We assume that the TTC threshold value is the summation 22.4% in B, and 10.3% in C. 30.6% of respondents who of the perception reaction time, the margin time for an LED sign, experienced accident risk indicated that those risks had and the vehicle stopping time. In this study, a fxed TTC occurred at night. Moreover, 26.5% were using a smart- threshold of four seconds is used to account for a safety margin phone, and 22.4% were wearing earphones (or headphones) to some extent. We consider the perception reaction time to be when the incidents happened. Regarding campus trafc 1.5 seconds [40], the LED sign margin time to be 1 second, and safety, 78.8% of respondents indicated that it should be the vehicle stopping time to be 1.5 seconds. improved for four reasons, as shown in Figure 11. First, there is no separation between streets for vehicles and sidewalks for pedestrians on campus roads (16.8%). Second, the road 2.4. Collision Risk Warning. If the situation is judged to be widths are narrow (15.1%). Tird, insufcient guiding signs dangerous, drivers and pedestrians are presented with LED sign on one-way roads frequently lead to wrong-way driving information. Tis service delivers warning information on the (7.6%). Fourth, there are multiple blind spots due to parked roadside for vehicles and pedestrians, as opposed to prior vehicles and buildings (5.9%). Tese factors are consistent systems that provided risk information only to vehicles. Te with the safety problems of residential roads in other regions warning information is presented in Figure 8. of Korea [6]. Ten, after service installation and operation on KAIST Seoul Campus, 76.9% of total subjects noted that the 3.Application and Evaluation service contributed to trafc safety on the campus from the 3.1. Application. We applied the proposed service on KAIST results of the Likert scale, as shown in Table 4. Tey stated Seoul Campus in Korea. Te service was provided in three that the service could prevent collision risk in the blind spots situations: illegal roadside parking, unprotected left turn, by providing warnings. Specifcally, they mentioned that and wrong-way driving. Figure 9 provides a description of LED sign boards made signs instantly recognizable, even at the application site and each situation. In addition, we night, compared to convex mirrors. In addition, 78.6%, evaluated the proposed service to analyze its impact on safety 77.6%, and 85.1% of respondents who were aware of the by conducting an on-site survey. service in Locations A, B, and C, respectively, believed that roads were safer after service operation. In addition, 92.9%, 91.8%, and 94.6% of respondents who knew the service in 3.2. Survey Design. After the service application, we analyzed Locations A, B, and C indicated that the service was nec- responses to the service in terms of safety efects. We collected essary for campus trafc safety, as presented in Figure 12. data through in-person interviews. Te survey questionnaire Second, we conducted a chi-square test to determine consisted of four sections. First, we inquired about demographic whether respondents’ perceptions of the service in operation characteristics, including gender, age, mobility impairment, and at three diferent locations difered signifcantly. A chi- current modes of transportation. Second, we asked whether square test is a nonparametric test to analyze the in- accidents or accident hazards had occurred at the site. If so, dependence or diference across a group among nominal respondents were questioned as to whether they were using variables [41]. We used the chi-square test of homogeneity to a smartphone or headphones and about locations of incidents. compare the proportions of service perception among Furthermore, we inquired whether respondents thought campus groups at three locations for signifcant diferences. To trafc safety should be improved. Tird, for each location, we conduct the homogeneity test, samples of the test groups solicited feedback on the installed and operating service, such as must be distinct [42]. For this, three diferent groups of preference or level of satisfaction. We frst inquired if re- respondents were questioned about their opinions of the spondents were aware that the service was operational. If they service in the three designated locations (i.e., Locations A, B, were familiar with the service, they were asked about its safety and C). Each location had a diferent number of individuals efects following the operation and its requirements at each who were aware that the service was operational. In Location location. Finally, the contribution of the service to safety on C, there were more individuals who were unaware that the KAIST Seoul Campus was assessed using a fve-point service was operational than those who were aware, as Likert scale. shown in Figure 12. Terefore, we were able to use the Te respondents to the questionnaire were people who homogeneity test. Te formula for the test statistics, χ commute to KAIST Seoul Campus. Te survey was conducted -value, is mathematically expressed as equation (8). in July 2022. A total of 151 responses were collected. Te L J majority of respondents were campus members, such as stu- 􏼐O − E 􏼑 l,j l,j χ � 􏽘 􏽘 , (8) dents, professors, and employees, while some were local resi- l,j l�1 j�1 dents and travelers who were passing by KAIST Seoul Campus. Sample descriptions are summarized in Table 4. where O is observed frequency, E is expected frequency, l l,j l,j represents the location index, j refers to the category of 3.3. Survey Analysis. We analyzed the collected data, in- response (e.g., yes or no), and L and J are the number of cluding descriptive statistical analysis and the chi-square locations and the number of categories, respectively. Te test. First, we investigated the perception of the trafc safety degree of freedom (df) was found to be two, using the Journal of Advanced Transportation 11 Figure 8: Collision risk warning as delivered by LED sign board. Tree diferent images are displayed sequentially until a risk situation is over. Korea Advanced Institute of Science and Technology (KAIST) Seoul Campus Illegal roadside parking Unprotected lef turn Wrong-way driving (Location A) (Location B) (Location C) Parked vehicle LED sign board Driving vehicle One-way Driving vehicle (Wrong-way) Pedestrian Figure 9: Site description: the proposed service was applied to three diferent locations on KAIST Seoul Campus. Each location was chosen for a single specifc risk circumstance: Location A for illegal roadside parking; Location B for unprotected left turn; and Location C for wrong-way driving. 12 Journal of Advanced Transportation Table 4: Sample description. Sample attributes Number of samples % Male 114 75.5 Gender Female 37 24.5 20’s 78 51.7 30’s 41 27.2 Age 40’s 15 9.9 50’s 12 7.9 Over 60 5 3.3 Yes 1 0.7 Mobility handicapped No 150 99.3 On foot 117 77.5 Vehicle 26 17.2 Transportation mode Bicycle 5 3.3 Personal mobility 3 2.0 Less than six months 52 34.4 Working or visiting period Six months∼one year 26 17.2 More than one year 73 48.4 Yes 49 32.5 Accident risk experience No 102 67.5 Necessary 119 78.8 Necessity to improve trafc safety on KAIST Seoul Campus Not necessary 32 21.2 (Strongly) did not contribute 2 1.3 Did not contribute 5 3.3 Contribution of service to trafc safety Neutral 28 18.5 Contributed 99 65.6 (Strongly) contributed 17 11.3 <Transportation mode> (63.3%) <Gender> <Using a smartphone> (28.6%) 34 13 (6.1%) (26.5%) (69.4%) (2.0%) On foot Vehicle <Collision risk experience> (30.6%) (73.5%) Bicycle Personal mobility <Location> Male Use a smartphone Female Not use a smartphone 49 11 (67.5%) (32.5%) (22.4%) <Age> <Wearing earphones> (44.9%) (10.3%) (28.6%) (22.4%) (14.3%) (22.4%) (38.8%) A C 38 Yes 7 (77.6%) (14.3%) B Other No <Time> (4.0%) Wear earphones 19~29 50~59 (69.4%) Not wear earphones 30~39 Over 60 40~49 (30.6%) Daytime Nighttime Figure 10: Responses to questions about collision risk. Respondents who experienced risk were asked seven diferent questions. Journal of Advanced Transportation 13 <Need for safety improvement> (15.1%) (7.6%) <Reasons> No separation between streets and sidewalks 32 119 (5.9%) (16.8%) Narrow road widths (21.2%) (78.8%) High frequency of the wrong-way driving (0%) Numerous blind spots Steep slope No response Yes No Figure 11: Responses to question about need for safety improvement. <Is it safer than before at Location A?> <Is it safer than before at Location B?> (78.6%) (77.6%) <Awareness of service at Location A> <Awareness of service at Location B> Safer than before Safer than before (35.1%) (43.7%) No diference No diference (56.3%) (64.9%) <Is the service necessary at Location A?> <Is the service necessary at Location B?> 91 78 Yes (92.9%) Yes (91.8%) No No 7 (7.1%) (8.2%) Necessary Necessary Not necessary Not necessary <Is it safer than before at Location C?> (85.1%) <Awareness of service at Location C> (14.9%) Safer than before 77 74 No diference (51.0%) (49.0%) <Is the service necessary at Location C?> Yes (94.6%) No (5.4%) Necessary Not necessary Figure 12: Perception of service at each location. following formula: Degree of freedom � (the number of We obtained two χ values of 1.6444 and 0.4911, with as- rows − 1) (the number of columns − 1). We established the sociated p values of 0.4395 and 0.7823, respectively. Since two diferent null hypotheses that the proportion among the these p values are greater than the signifcance level of 0.05, three groups is the same for the following questions: (i) “Do we do not reject the two hypotheses. Te report states that you think that the service has increased safety at this lo- there is no statistically signifcant diference among the cation?” and (ii) “Do you think that the service is necessary locations in terms of service perception. Each location at this location?.” If the calculated χ -value is greater than the represents scenarios that can happen on residential roads. critical value from the χ -distribution, we must reject the Accordingly, it is demonstrated that the proposed service null hypothesis. Tis implies that at least one proportion can be implemented in many locations on residential roads difers considerably from another proportion among groups. and have the same efect regardless of location from a user’s 14 Journal of Advanced Transportation Table 5: Chi-square test results. Number of responses (%) Results Question Location Yes No Total χ df p value A 77 (78.6%) 21 (21.4%) 98 (1) Do you think that the service has increased safety at this location? B 66 (77.6%) 19 (22.4%) 85 1.6444 2 0.4395 C 63 (85.1%) 11 (14.9%) 74 A 91 (92.9%) 7 (7.1%) 98 (2) Do you think that the service is necessary at this location? B 78 (91.8%) 7 (8.2%) 85 0.4911 2 0.7823 C 70 (94.6%) 4 (5.4%) 74 Table 6: Suggestions for improving trafc safety at each location, based on open-ended questions. Ideas for Ideas for Ideas for % % % Location A Location B Location C Use acoustic speakers 20 Need advertisements 19 Use acoustic speakers 25 Need advertisements 16 Use acoustic speakers 11 Need trafc safety signs 22 Need convex mirrors 13 Prohibit illegal roadside parking 11 Install automatic roadblock 14 perspective. Te results of the chi-square test are summa- improving trafc safety on KAIST Seoul Campus. In ad- rized in Table 5. dition, they stated that the campus was safer than before the Finally, we asked respondents for suggestions on im- service installation, with 78.6%, 77.6%, and 85.1% proving trafc safety at each location. Te suggestions are responding positively for Locations A, B, and C, respectively. summarized in Table 6. One of the most recommended Tese results, combined with the high satisfaction reported approaches at almost all locations was to use an acoustic by survey respondents, suggest that our service can be ap- speaker to deliver warning information. Some respondents plied to various areas that are typically considered residential suggested that it would be more efective if both a visual and roads. Te service implementation is expected to improve an audible warning were utilized concurrently. Another trafc safety and reduce fatalities that arise due to blind spots. suggestion was to promote the service to campus members. Indeed, in Locations A, B, and C, 64.9%, 56.3%, and 49.0% of Te generalizability of these results is subject to certain respondents, respectively, knew that the service was in limitations. For instance, evaluating the proposed service operation. Except for the use of acoustic speakers, Location was conducted based only on a survey. A natural progression C’s suggestions difered from those of the other locations. of this work is to evaluate the service using before-and-after Te characteristics of Location C included trafc safety signs surrogate data, such as the frequency of two- and three- for one-way driving and an automatic roadblock that pre- second TTC events. We are currently collecting the relevant vents vehicles from traveling in the incorrect direction. data for this purpose. In addition, TTC may also vary based on given infrastructure, nature, and human environments. In this study, a constant TTC threshold was adopted with 4.Conclusions a safety margin, which could yield unnecessary false posi- tives in some situations. Terefore, we should conduct In this study, we propose a collision risk warning service for a sensitivity analysis for TTC with respect to diferent residential roads based on risk assessment. In contrast to location-specifc environments to determine the optimal earlier research, this service combines CCTVs and radar to TTC for each location. Moreover, using acoustic speakers to detect items precisely and quickly. We use an elliptical alert vehicles and pedestrians appeared to be the most trajectory prediction approach to predict unknown pedes- suggested approach for enhancing trafc safety at all loca- trian behaviors. Te major and minor axes of the ellipse were tions. According to the respondents, they believed that derived using CCTV data on actual pedestrian trajectories. combining visual and audible warnings would provide Furthermore, we use TTC to identify collision risks in a more efective warning to those using a smartphone and/or vehicle-vehicle and vehicle-pedestrian cases. An LED sign earphones. Te results are consistent with [43], which board is used to provide risk warnings to vehicles and pe- supported the idea that multimodal warning services have destrians. Te proposed service was provided in three sit- potential advantages in various situations, such as when uations: illegal roadside parking, unprotected left turn, and people are using smartphones or are engaged in distracted wrong-way driving on residential roads. driving. We applied our service to three locations on KAIST Seoul Campus in Korea. To evaluate service efects, we conducted a survey and analyzed the safety efects from the Data Availability user’s perspective. Using a set of questions, we investigated respondents’ satisfaction with the service. 76.9% of re- Te data used to support the fndings of this study are spondents reported that the service contributed to available from the corresponding author upon request. Journal of Advanced Transportation 15 [13] B. Noh and H. Yeo, “A novel method of predictive collision Conflicts of Interest risk area estimation for proactive pedestrian accident pre- vention system in urban surveillance infrastructure,” Trans- Te authors declare that they have no conficts of interest. portation Research Part C: Emerging Technologies, vol. 137, Article ID 103570, 2022. Authors’ Contributions [14] W. Yang, X. Zhang, Q. Lei, and X. Cheng, “Research on longitudinal active collision avoidance of autonomous Aya Selmoune and Jeongin Yun contributed equally to emergency braking pedestrian system (AEB-P),” Sensors, this paper. vol. 19, no. 21, 2019. [15] R. Matsumi, P. Raksincharoensak, and M. Nagai, “Autono- Acknowledgments mous braking control system for pedestrian collision avoid- ance by using potential feld,” IFAC Proceedings Volumes, Tis work was supported by the Ministry of the Interior and vol. 46, no. 21, pp. 328–334, 2013. Safety (MOIS), Republic of Korea (grant nos. 2021-MOIS41- [16] A. Høye and A. Laureshyn, “SeeMe at the crosswalk: before- 001-00000000-2022 (development of trafc safety risk after study of a pedestrian crosswalk warning system,” Transportation Research Part F: Trafc Psychology and Be- warning technology for blind spots in living roads)). haviour, vol. 60, pp. 723–733, 2019. [17] D. Ka, D. Lee, and H. Yeo, “Development of predictive pe- References destrian collision warning service considering pedestrian characteristics,” Te Journal of Te Korea Institute of In- [1] WHO, “WHO road trafc injuries,” 2022, https://www.who. telligent Transport Systems, vol. 18, no. 3, pp. 68–83, 2019. int/news-room/fact-sheets/detail/road-trafc-injuries. [18] J. J. Anaya, P. Merdrignac, O. Shagdar, F. Nashashibi, and [2] PWC, “A guide for policy makers: on reducing road fatalities,” J. E. Naranjo, “Vehicle to pedestrian communications for 2022, https://www.pwc.com/m1/en/publications/guide-on- protection of vulnerable road users,” in Proceedings of the reducing-road-fatalities.html. IEEE Intelligent Vehicles Symposium Proceedings, Dearborn, [3] “OECD International Transport Forum: Road injury acci- MI, USA, June 2014. dents,” 2022, https://stats.oecd.org/Index.aspx?DataSetCode= [19] K. Dhondge, S. Song, B. Choi, and H. Park, “WiFiHonk: ITF_ROAD_ACCIDENTS. smartphone-based beacon stufed WiFi Car2X- [4] M. W. Adler and R. Ahrend, “Trafc safety in South Korea: communication system for vulnerable road user safety,” in understanding the vulnerability of elderly pedestrians,” OECD Proceedings of the IEEE 79th Vehicular Technology Conference Regional Development Working Papers, Trafc Safety in Korea, (VTC Spring), pp. 1–5, Seoul, Korea, May 2014. OECD Publishing, Seoul, Korea, 2017. [20] T. Wang, G. Cardone, A. Corradi, L. Torresani, and [5] “OECD International Transport Forum: Road safety annual A. T. Campbell, “Walksafe: a pedestrian safety app for mobile report 2021 korea,” 2022, https://www.itf-oecd.org/sites/ phone users who walk and talk while crossing roads,” in default/fles/korea-road-safety.pdf. Proceedings of the Twelfth Workshop on Mobile Computing [6] J. Lim, S. Lee, J. Choi, and S. Joo, “Te comparative study on Systems & Applications, pp. 1–6, San Diego, CA, USA, Feb- travel behavior and trafc accident characteristics on a com- ruary 2012. munity road-with focus on Seoul metropolitan city,” Journal [21] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only of the Korean Society of Civil Engineers, vol. 36, no. 1, Look once: unifed, real-time object detection,” in Proceedings pp. 97–104, 2016. of the IEEE Conference on Computer Vision and Pattern [7] S. Tak, S. Kim, D. Lee, and H. Yeo, “A comparison analysis of Recognition, pp. 779–788, las vegas, NV, USA, June 2016. surrogate safety measures with car-following perspectives for [22] F. Zhou, H. Zhao, and Z. Nie, “Safety helmet detection based advanced driver assistance system,” Journal of Advanced on YOLOv5,” in Proceedings of the IEEE International Con- Transportation, vol. 2018, Article ID 8040815, 14 pages, 2018. ference on Power Electronics, Computer Applications [8] S. o. Son, J. Park, C. Oh, and C. Yeom, “An algorithm for (ICPECA), pp. 6–11, Shenyang, China, January 2021. detecting collision risk between trucks and pedestrians in the [23] R. Padilla, S. L. Netto, and E. A. Da Silva, “A survey on connected environment,” Journal of Advanced Trans- performance metrics for object-detection algorithms,” in portation, vol. 2021, Article ID 9907698, 9 pages, 2021. Proceedings of the 2020 international conference on systems, [9] L. Zhang, K. Yuan, H. Chu et al., “Pedestrian collision risk signals and image processing (IWSSIP), pp. 237–242, IEEE, assessment based on state estimation and motion prediction,” Niteroi, Brazil, July 2020. IEEE Transactions on Vehicular Technology, vol. 71, no. 1, [24] R. Solovyev, W. Wang, and T. Gabruseva, “Weighted boxes pp. 98–111, 2022. fusion: ensembling boxes from diferent object detection [10] Y. Wu, M. Abdel-Aty, O. Zheng, Q. Cai, and L. Yue, “De- models,” Image and Vision Computing, vol. 107, Article ID veloping A crash warning system for the bike lane area at 104117, 2021. intersections with connected vehicle technology,” Trans- [25] I. Culjak, D. Abram, T. Pribanic, H. Dzapo, and M. Cifrek, “A portation Research Record, vol. 2673, no. 4, p. 47, 2019. [11] S. Zhang, M. Abdel-Aty, Y. Wu, and O. Zheng, “Modeling brief introduction to OpenCV,” in Proceedings of the 35th International convention MIPRO, pp. 1725–1730, Opatija, pedestrians’ near-accident events at signalized intersections using gated recurrent unit (GRU),” Accident Analysis & Croatia, May 2012. [26] B. Noh, W. No, and D. Lee, “Vision-based overhead front Prevention, vol. 148, Article ID 105844, 2020. [12] S. Zhang, M. Abdel-Aty, Y. Wu, O. Zheng, and J. Ugan, point recognition of vehicles for trafc safety analysis,” in Proceedings of the 2018 ACM International Joint Conference “Prediction of pedestrian-vehicle conficts at signalized in- tersections based on long short-term memory neural net- and 2018 International Symposium on Pervasive and Ubiq- uitous Computing and Wearable Computers, pp. 1096–1102, work,” Accident Analysis & Prevention, vol. 148, Article ID 105799, 2020. Singapore, October 2018. 16 Journal of Advanced Transportation [27] S. S. Pathan, A. Al-Hamadi, and B. Michaelis, “Intelligent [42] T. M. Franke, T. Ho, and C. A. Christie, “Te chi-square test: feature-guided multi-object tracking using kalman flter,” in often used and more often misinterpreted,” American Journal of Evaluation, vol. 33, no. 3, pp. 448–458, 2012. Proceedings of the 2nd International Conference on Computer, [43] F. Biondi, D. L. Strayer, R. Rossi, M. Gastaldi, and C. Mulatti, Control and Communication, pp. 1–6, Karachi, Pakistan, “Advanced driver assistance systems: using multimodal re- February 2009. dundant warnings to enhance road safety,” Applied Ergo- [28] X. Li, K. Wang, W. Wang, and Y. Li, “A multiple object nomics, vol. 58, pp. 238–244, 2017. tracking method using kalman flter,” in Proceedings of the IEEE International Conference on Information and Automa- tion, pp. 1862–1866, Harbin, China, June 2010. [29] N. S. Punn, S. K. Sonbhadra, S. Agarwal, and G. Rai, “Monitoring COVID-19 social distancing with person de- tection and tracking via fne-tuned YOLO v3 and Deepsort techniques,” 2020, https://arxiv.org/abs/2005.01385. [30] A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in Proceedings of the 2016 IEEE international conference on image processing (ICIP) 2016, pp. 3464–3468, Phoenix, AZ, USA, September 2012. [31] G. Shu, A. Dehghan, O. Oreifej, E. Hand, and M. Shah, “Part- based multiple-person tracking with partial occlusion han- dling,” in Proceedings of the 2012 IEEE Conference on Com- puter Vision and Pattern Recognition 2012, pp. 1815–1821, Rhode Island, RI, USA, June 2012. [32] K. Bernardin and R. Stiefelhagen, “Evaluating multiple object tracking performance: the clear mot metrics,” EURASIP Journal on Image and Video Processing, vol. 2008, Article ID 246309, 10 pages, 2008. [33] K. S. Oh, S. Y. Park, J. H. Seo, G. H. Lee, and K. S. Yi, “Laser- Scanner-based stochastic and predictive working-risk- assessment algorithm for excavators,” Journal of Drive and Control, vol. 13, no. 4, pp. 14–22, 2016. [34] M. Seo, H. Shin, H. Jeong, and J. Chae, “Development of an object collision detection algorithm for prevention of collision accidents on living roads,” Journal of Drive and Control, vol. 19, no. 3, pp. 23–31, 2022. [35] R. Miller and Q. Huang, “An adaptive peer-to-peer collision warning system Vehicular Technology Conference,”vol. 1, pp. 317–321, in Proceedings of the IEEE 55th Vehicular Technology Conference. VTC Spring, vol. 1, pp. 317–321, IEEE, Birmingham, AL, USA, May 2020. [36] F. Jimenez, ´ J. E. Naranjo, and F. Garc´ıa, “An improved method to calculate the time-to-collision of two vehicles,” International Journal of Intelligent Transportation Systems Research, vol. 11, no. 1, pp. 34–42, 2013. [37] R. Van Der Horst and J. Hogema, Time-to-collision and Collision Avoidance Systems, in Proceedings of the 6th ICTCT Workshop Salzburgna, Salzburg, Te Netherlands, 1993. [38] S. Das and A. K. Maurya, “Defning time-to-collision thresholds by the type of lead vehicle in non-lane-based trafc environments,” IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 12, pp. 4972–4982, 2020. [39] F. Bella and R. Russo, “A collision warning system for rear- end collision: a driving simulator study,” Procedia-social and behavioral sciences, vol. 20, pp. 676–686, 2011. [40] U. Durrani, C. Lee, and D. Shah, “Predicting driver reaction time and deceleration: comparison of perception-reaction thresholds and evidence accumulation framework,” Acci- dent Analysis & Prevention, vol. 149, 2021. [41] M. L. McHugh, “Te chi-square test of independence,” Bio- chemia Medica, vol. 23, no. 2, pp. 143–149, 2013.

Journal

Journal of Advanced Transportation – Hindawi Publishing Corporation

Published: Mar 16, 2023

There are no references for this article.