Open Advanced Search
Get 20M+ Full-Text Papers For Less Than $1.50/day.
Start a 14-Day Trial for You or Your Team.
Learn More →
Target Detection of Low-Altitude UAV Based on Improved YOLOv3 Network
Target Detection of Low-Altitude UAV Based on Improved YOLOv3 Network
Zhai, Haiqing;Zhang, Yang
Hindawi Journal of Robotics Volume 2022, Article ID 4065734, 8 pages https://doi.org/10.1155/2022/4065734 Research Article Target Detection of Low-Altitude UAV Based on Improved YOLOv3 Network 1,2 1,2 Haiqing Zhai and Yang Zhang School of Computer Science and Technology, Henan Institute of Technology, Xinxiang, Henan 453003, China Big Data Engineering Research Center of Henan for Production and Manufacturing IoTs, Xinxiang, Henan 453003, China Correspondence should be addressed to Haiqing Zhai; email@example.com Received 7 January 2022; Accepted 18 February 2022; Published 10 March 2022 Academic Editor: Shan Zhong Copyright © 2022 Haiqing Zhai and Yang Zhang. *is 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. Most existing methods are diﬃcult to detect low-altitude and fast-moving drones. A low-altitude unmanned aerial vehicle (UAV) target detection method based on an improved YOLOv3 network is proposed. While keeping the basic framework of the original model unchanged, the YOLOv3 model is improved. *at is, multiscale prediction is added to enhance the detection ability of small-target objects. In addition, the two-axis Pan/Tilt/Zoom (PTZ) camera is controlled based on proportional integral derivative (PID), so that the target tends to the center of the ﬁeld of view. It is more conducive to accurate detection. Finally, experiments are carried out using real UAV datasets. *e results show that the mean average precision (mAP), AP50, and AP75 are 25.12%, −1 39.75%, and 26.03%, respectively, which are better than other methods. Also, the frame rate is 21 frames·s , which meets the performance requirements. susceptible to interference from external environmental 1. Introduction factors. *e method based on image recognition is limited by With the rise and development of UAV technology, it has the bottleneck of computer technology and communication been widely used in military and civil ﬁelds. However, a large technology and has not been widely used and developed number of UAVs pose a certain threat to the ﬂight safety of [6, 7]. At present, there has been a certain amount of re- aircraft and the political sensitivity of images in conﬁdential search on UAV target recognition in China. For example, areas. At the same time, it also brings huge challenges to *illainayagi and Senthil Kumar  proposed a target de- urban security . For the sake of public safety, local tection technology for UAV thermal images based on governments prohibit unauthorized UAV ﬂight in airports, wavelet transform and singular value decomposition, using meeting places, and other areas . *erefore, monitoring discrete wavelet transform and stationary wavelet transform to enhance image texture features and edge features. *e UAVs in speciﬁc areas is an urgent need for security. Due to the small volume and low speed of UAV, it is diﬃcult to experimental results show that the proposed method has detect UAV by using traditional radar equipment . In smaller errors. However, the detection eﬃciency needs to be addition, the noise of the city is noisy, so it is diﬃcult to improved. Li et al.  proposed a fast and eﬀective moving detect UAVs by acoustic sensors . *erefore, how to target detection method, which extracts cross features based detect UAV eﬃciently and accurately from the environment on line segments. It has good detection speed and rotation has become an urgent problem to be solved. accuracy, but the application scope of the algorithm is small So far, the identiﬁcation methods of UAVs are diverse. and the practical application has great limitations. Abdul- *e main methods focus on image recognition and radar ridha et al.  proposed a UAV hyperspectral image data analysis . However, traditional analysis methods that recognition method. Multilayer perceptual neural network rely on spectrum detection and radar data are extremely and stepwise discriminant analysis are used to realize the 2 Journal of Robotics Type ﬁlters size output image detection, which eﬀectively improves the detection Convolutional 32 3*3 256*256 accuracy. Yang et al.  proposed a UAV object detection Convolutional 64 3*3/2 128*128 model based on rotation and constant depth denoising in Convolutional 32 1*1 view of the diﬃculties of the multidirectional object, small 1× Convolutional 64 3*3 pixel, and vibration interference of the UAV body in the Residual 128*128 process of aerial object detection. A selective search method Convolutional 128 3*3/2 64*64 is used to extract the region of interest in the aerial image, Convolutional 64 1*1 and the radial gradient of the region of interest is calculated. 2× Convolutional 128 3*3 Combined with the deep denoising autoencoder, the original Residual 64*64 data noise is ﬁltered out and the deep features are extracted Convolutional 32 3*3 256*256 to realize the detection of aerial image targets. However, deep-level feature extraction will increase the amount of Convolutional 128 1*1 8× Convolutional 256 3*3 calculation and aﬀect the detection speed of the algorithm. Residual 32*32 In addition, image-based target detection technology has gradually been widely used with the signiﬁcant improvement Convolutional 512 3*3/2 16*16 of computing power and communication technology . Convolutional 256 1*1 Xiaofei  proposes a UAV multitarget tracking and path 8× Convolutional 512 3*3 planning method combining basic gray wolf optimizer and Residual 16*16 Gaussian distribution estimation. It overcomes the problem Convolutional 1024 3*3/2 8*8 of real-time optimization of complex projects with traditional Convolutional 512 1*1 models and has good eﬀectiveness and practicability. How- 4× Convolutional 1024 3*3 ever, the algorithm focuses on target tracking and path Residual 8*8 planning, and its performance in target detection needs to be Avgpool global improved. Tao et al.  proposed a target search strategy for Connected 1000 UAV based on reinforcement learning. *rough the rein- Somax forcement learning training, the image captured by the drone Figure 1: Network structure of YOLOv3 algorithm. is analyzed and processed to achieve target detection and tracking, but the detection eﬀect of maneuvering targets is pyramid network (FPN) and uses multiscale features for poor. Liu and Zhang  proposed an automatic vehicle target detection. On the premise of maintaining the speed detection method based on deep learning. Based on the in- advantage, the detection accuracy is further improved, es- teractive multimodel particle ﬁlter algorithm, the perfor- pecially the detection ability for small targets is strength- mance of the UAV for maneuvering target positioning has ened. Also, the detection eﬀect of high-coverage images is been signiﬁcantly improved. However, the detection eﬀect of signiﬁcantly higher than that of YOLOv2 . small moving targets such as drones has yet to be veriﬁed. *e YOLO series of algorithms constantly iteratively Based on the above analysis, a target detection algorithm optimize the detection accuracy on the basis of always based on the improved YOLOv3 network is proposed for the maintaining the advantage of high detection speed and problem of low-altitude UAV target detection. *e inno- gradually improve the detection ability of small targets. At vations are summarized as follows: the same time, accurate detection of high-coverage images is (1) In order to make up for the inability of the YOLOv3 realized, the structure is simple, and the background false network to detect small targets, the proposed method detection rate is low. It has become the most popular real- adds multitarget prediction to improve the YOLOv3 time target detection algorithm today. network. Four diﬀerent sizes of bounding boxes are provided to match the actual bounding box as much 2.2. Improved YOLOv3. In the detection and recognition as possible, thereby improving the accuracy of target process, when two or more types of similar objects appear, detection. the YOLO network will often misidentify them as the same (2) Since drone monitoring is costly and diﬃcult to type. *is shows that the YOLO network is too poor to implement, the proposed method uses a two-axis distinguish the details. At the same time, the recognition PTZ camera to track the drone. Among them, the performance of YOLO for objects with large diﬀerence in PID algorithm is used to adjust the camera position length and width ratio needs to be improved . When the to achieve eﬃcient detection of low-altitude UAVs. proportion of large and small objects in the image diﬀers greatly, the large objects can be identiﬁed in the test results, but the small objects often cannot be detected [18, 19]. *e 2. Target Detection Based on Improved YOLOv3 UAV remote sensing image used in the experiment has a 2.1. YOLOv3 Network Structure. *e YOLOv3 algorithm lower resolution, a larger image scale, rich information, and uses a fully convolutional network composed of residual more complex details. Direct use of conventional network blocks as the backbone network. *e network depth reaches models for detection and recognition is often not ideal 53 layers and is called Darknet-53. Its network structure is [20, 21]. *erefore, the model algorithm needs to be further shown as in Figure 1. YOLOv3 draws on the idea of Feature improved according to the characteristics of UAV images. Journal of Robotics 3 For the above reasons and the analysis of the YOLO model, the YOLOv3 model is improved while maintaining the basic framework of the original model, that is, multiscale prediction is added. In the YOLOv2 network model, in order to enhance the accuracy of small-target detection, the feature map extracted Image Conv Conv Conv from the last layer of the network model is connected with the feature map of the previous layer through the pass- through layer. *e size of the feature map of the last layer is Multi-scale 13×13. In YOLOv3, this method is further enhanced. YOLOv3 provides three bounding boxes of diﬀerent sizes, using similar concepts to extract features of these sizes to form a pyramidal network. YOLOv3 adds several con- Precision volutional layers, and the ﬁnal convolution layer is used to predict the tensor coding including boundary box, target in Figure 2: Multiscale prediction. the box, and classiﬁcation prediction . YOLOv3 uses the feature fusion of multiple scales, so the number of bounding boxes is much more than before. Input image Adding a scale prediction to the YOLOv3 network model provides four bounding boxes of diﬀerent sizes, as shown in Darknet-53 feature extraction Figure 2. *e improved YOLOv3 network includes the last network layer, and the feature map is 13×13. *ere are also 3 upsampled eltwise sums with feature maps of 26×26, 52×52, and 104×104. *e largest network model uses the 104 feature Feature pyramid map. Also, YOLOv2 takes multiscale into consideration for the training data sampling, and in the end, only the feature Detection Detection Detection map of 13 is used. *is should be the place that has the layer layer layer greatest impact on small goals . At the same time, the input image size is changed to 1024×1024 to adapt to large- Multiscale detection process scale sampling. In the experiment, 12 clusters were selected. *en, divide the dimensional clusters evenly on bounding Non maximum suppression boxes of diﬀerent sizes. *e 12 clusters are (10×13), (16×30), (33×23), (30×61), (62×45), (59×119), (116×90), (156 × 198), Figure 3: Detection process of YOLOv3 algorithm. (373 × 326), (312 × 536), (584 × 712), and (869 × 796). p � sigmoid t + b , x x x 2.3. Detection Process. *e improved YOLOv3 algorithm does not need to generate a region of interest (ROI) in p � sigmoidt + b , y y y (2) advance but directly trains the network in a regression way. p � g × e , At the same time, the k-means algorithm is used to cluster w w the sample bounding boxes, and four groups of bounding h p � g × e , h h box sizes are preset on the four scale sizes respectively, so as to make positioning prediction based on the bounding where b and b represent the oﬀset of the grid to which x y boxes of 16 sizes. *e whole detection process is shown in the bounding box belongs relative to the upper left corner Figure 3. of the image. g and g represent the length and width of h w First, feature extraction is performed on the original the predeﬁned bounding box. p and p indicate the x y 1024×1024 input image through the feature extraction distance from the center of the bounding box of the ﬁnal network. *en, the feature vector is fed into the prediction result to the upper left corner of the image. p FPN structure to generate grid areas on 4 scales. *ey are and p are the length and width of the predicted bounding 13×13, 26×26, 52×52, and 104×104. Each grid area predicts 4 box. bounding boxes, resulting in a total of *e l of the vector Ω is expressed as follows: (104×104+52×52+26×26+13×13)×4 �57460 bounding gt l � Θ(o) × IoU , (3) 0 o boxes. Next, a vector Ω is predicted in each bounding box, which is expressed as follows: where Θ(o) represents the probability that the object is in gt the prediction frame. IoU represents the intersection over Ω � t + t + t + t + l + l + l + · · · + l . (1) x y w h 0 1 2 n union (IoU) of the predicted box and the true bounding box. *e ﬁrst 4 elements t , t , t , and t in the vectorΩ are the When using logistic regression to score the prediction box x y w h 4 coordinates related to the bounding box, and their rela- the highest, the probability that the object is in the prediction tionship is as follows: box is 1, otherwise it is 0. *e l in vector Ω represents the 0 4 Journal of Robotics score that the predicted object belongs to one of the classes. When the prediction frame is obtained, the nonmaximum suppression is carried out to obtain the ﬁnal prediction result. 3. Design of UAV Vision following Control Algorithm *e two-axis PTZ camera is shown in Figure 4. *e function of the camera is to control the movement of the camera to keep the target in the center of the video. *e control module is a two-axis PTZ with two steering gears. One servo is responsible for controlling the camera to move left and right, and the other controls the camera to move up and down. Each steering gear has an adjustment range of 180 . *e PID control algorithm is expressed as follows: dθ(t) ′ ′ q(t) � κ θ(t) + κ θ t dt + κ , (4) p i d Figure 4: Two-axis PTZ camera. dt where q(t) is the output of the system, which means the steering gear rotation angle, rad; θ(t) is the deviation angle Camera between the image center and the UAV center, rad; κ , κ , p i and κ are all constant coeﬃcients, corresponding to pro- portional gain, integral gain, and diﬀerential gain [24, 25]. Next Equation (4) is composed of three parts: proportional, Frame integral, and diﬀerential. *e ﬁrst part makes the camera rotate with the movement of the drone. *e integral part is used to eliminate the stabilization error and prevent the drone from being out of the video center. *e derivative part Improved YOLOv3 is used to control the rate of change of the deviation . *e PID control process is shown in Figure 5. Using Failure Success OpenCV (Computer Vision Library) to process the camera video stream, improve YOLOv3 detection for each frame of Object No object the video stream. After obtaining the position of the drone in coordinate the picture, calculate the distance between its center and the center of the picture. *e distance parameter is passed to the PID process for calculation, so as to control the rotation of PID the steering gear. (Pan and tilt) Figure 5: PID control process. 4. Experiment and Analysis *e experimental platform is a computer with Intel core i7- rapid ascent and descent, and smooth ﬂight, were fully 7700HQ2.8GHz CPU and GeForce 1050 ti 2GB GPU. *e considered during the shooting process. In the end, 3864 minibatch during training is set to 5, and the learning rate is visible light images were obtained. 0.001. 4.1.2. Data Annotation. In order to ensure the validity of the 4.1. Dataset. *ere are few low-altitude UAV detection and data, the manual labeling method is adopted to label the recognition methods based on deep learning, and there is no samples whose target occluded area is greater than or equal public dataset or standard dataset. *erefore, ﬁrstly, the to 50%. Finally, the bounding boxes of the low-altitude dataset is collected and constructed. drone targets in the 3258 visible light images were labeled, and data with labeled information were obtained. According 4.1.1. Data Collection. Using visible light detectors and two- to the ratio of 5:1, it is divided into the training set and test axis pan-tilt cameras to take images of 4 types of civilian set. drones at diﬀerent times and in diﬀerent backgrounds, the UAV models are DJI-Elf 3 (DJ-3), DJI-Yu Pro (DJ-Pro), DJI- Yu Mavic 2 zoom version (DJ-M2 Z), and DJI-Yu Air (DJ- 4.1.3. Image Enhancement. In order to improve the detec- Air). In order to ensure the diversity of data, the various tion and recognition accuracy of the method, the general ﬂight attitudes of low-altitude drones, including hovering, image enhancement method in the ﬁeld of target detection is Journal of Robotics 5 adopted to enhance the training set, including the operation detection accuracy of DJ-3 and DJ-M2 Z are 91.74% and processing of brightness and contrast. 91.98%, respectively. *e image performance characteristics of DJ-3 and DJ-M2 Z are similar, so they are easy to confuse. According to Tables 2 and 3, the classiﬁcation eﬀect of 4.1.4. Data Expansion. Taking into account that the attitude the YOLOv3 algorithm and the improved YOLOv3 algo- of the UAV during ﬂight is not completely horizontal and rithm on the dataset are obtained, as shown in Table 4. inclined, etc., the training set is ﬂipped and rotated at ±10 It can be seen from Table 4 that the improved YOLOv3 and ±20 . If the target near the edge in the image is damaged algorithm has a certain improvement in the recognition and or completely lost after the rotation processing, the sample detection eﬀect compared with the classic YOLOv3 algo- data are discarded. rithm. Compared with the YOLOv3 algorithm, the average detection accuracy rate is increased by about 1.5%. *e 4.1.5. Dataset Construction. *rough the enhancement and improved YOLOv3 algorithm adds multiscale prediction, expansion of the images in the training set, three types of which can strengthen the detection of small targets. *erefore, DJ-Pro, DJ-M2 Z, etc., can be better datasets are obtained. As shown in Table 1, the training set of UAV-A is composed of low-altitude UAV targets in the distinguished. original image. *e training set of UAV-B is composed of the original image (UAV-A) and the image after image enhancement processing. *e training set of UAV-C is 4.4. Performance Comparison with Comparison Algorithm. composed of UAV-A and UAV-B and their expanded In order to demonstrate the performance of the proposed samples (including data sets UAV-A and UAV-B). Among method, comparison with [8, 11], and  was performed. them, the test set uses the same sample for method Among them, the comparison experiment stage evaluates veriﬁcation. the models obtained by diﬀerent mainstream target detec- tion methods through the same data training. In the ex- 50 75 periment, the average accuracy mAP, AP , and AP , and 4.2. Visual Control Field Test Experiment. *e target UAV frame rate evaluation indicators were used to perform hovers at the center of the ﬁeld of view of the cooperative quantitative analysis of detection accuracy and detection UAV at a relatively long initial distance to make speed. *e results are shown in Table 5. Among them, AP ′ ′ x (t) � x ,Δ(t)<Δ . In order to quantitatively measure the is an eﬀective index to evaluate the classiﬁcation ability of the size error between the actual bounding box and the expected algorithm. *e AP can reﬂect the ability of the detection bounding box, the bounding box size error is deﬁned as frame to return to the position of the bounding box. ϑ (t) � (Δ(t) − Δ)/Δ. *e experimental result of step re- It can be seen from Table 5 that the proposed method has sponse is shown in Figure 6. 50 75 been greatly improved, and its mAP, AP , and AP are It can be seen from Figure 6 that when t �2.5s, the v(t) 25.12%, 39.75%, and 26.03%, respectively. *is shows that step response curve quickly stabilizes. However, there is still the proposed improved YOLOv3 target detection frame- a certain steady-state error between the actual bounding box work shows better classiﬁcation ability and higher frame and the expected bounding box. regression accuracy. *e main reason for the obvious im- provement of detection accuracy is the use of multilevel 4.3. Comparison with YOLOv3 Algorithm Classiﬁcation feature maps with diﬀerent scales for target prediction. *is Eﬀect. In order to describe the classiﬁcation ability of the greatly improves the detection eﬀect of various targets that improved YOLOv3 algorithm for the dataset, the classiﬁ- change with the drone’s viewing angle and ﬂying height. In cation confusion matrix on the data set is calculated. *e addition, multiscale target prediction can predict the posi- YOLOv3 classiﬁcation confusion matrix and the improved tion and shape of the candidate frame based on image YOLOv3 classiﬁcation confusion matrix are shown in Ta- features and generates sparse and arbitrary-shaped candi- bles 2 and 3, respectively. date frames, which more closely match the real target frame. It can be seen from Tables 2 and 3 that DJ-Air has the In addition, the two-axis PTZ camera is controlled based on best detection eﬀect. *e accuracy obtained by YOLOv3 and PID, so that the target tends to the center of the ﬁeld of view, improved YOLOv3 algorithm reached 92.74% and 93.26%, which is more conducive to target recognition. *illainayagi respectively. Compared with other types of UAVs, DJ-Air and Senthil Kumar  realize target detection in UAV has obvious characteristics, irregular shape, and easy to thermal image based on wavelet transform and singular distinguish. *e detection accuracy of YOLOv3 algorithm value decomposition. *e detection model is simple and easy for DJ-3, DJ-Pro, and DJ-M2 Z is 88.69%%, 92.41%, and to implement, but the detection accuracy in a complex 89.91%, respectively. *e colors of DJ-3, DJ-Pro, and DJ-M2 environment is not high. Its mAP is only 16.14%. Yang et al. Z are not obvious, and the characteristics have certain  proposed a UAV object detection model based on similarities. In addition, the DJ-Pro and DJ-Air detection rotation invariant depth denoising. A deep denoising results are better in the modiﬁed YOLOv3 detection results, autoencoder is used to ﬁlter out the noise of the original data with the accuracy reaching 92.51% and 93.26%. Mainly and extract the deep features to realize the target detection of because these categories are quite diﬀerent from categories aerial images. Compared with reference , its detection other than themselves. Compared with other types of UAVs, accuracy has been improved. However, it is diﬃcult to the target features are obvious and easier to distinguish. *e accurately detect small target objects, and it is easy to cause 6 Journal of Robotics Table 1: Processing methods and sample numbers of diﬀerent training sets. Train dataset Processing method Number of samples UAV-A Original 3200 UAV-B Original+image enhancement 6400 UAV-C Original+image enhancement+data augmentation 65000 0.20 0.30 0.25 0.15 0.20 0.15 0.10 0.10 0.05 0.05 0.00 0.00 -0.05 -0.05 -0.10 012345678 012345678 Time (s) Time (s) (a) (b) Figure 6: Experimental results of step response. (a) Step response curve of ﬂight speed. (b) Boundary frame size error variation curve. Table 2: Confusion matrix of YOLOv3 classiﬁcation results. Category DJ-3 DJ-Pro DJ-M2 Z DJ-Air DJ-3 6382 49 187 21 DJ-pro 274 2336 141 39 DJ-M2 Z 401 108 3625 67 DJ-air 139 35 79 1239 Accuracy rate, % 88.69 92.41 89.91 92.74 Table 3: Confusion matrix of improved YOLOv3 classiﬁcation results. Category DJ-3 DJ-Pro DJ-M2 Z DJ-Air DJ-3 6701 43 102 10 DJ-Pro 236 2286 117 26 DJ-M2 Z 295 107 3306 59 DJ-Air 72 35 69 1314 Accuracy rate, % 91.74 92.51 91.98 93.26 Table 4: Classiﬁcation accuracy of two algorithms on datasets. Category YOLOv3 Improved YOLOv3 DJ-3 88.69 91.74 DJ-Pro 92.41 92.51 DJ-M2 Z 89.91 91.98 DJ-Air 92.74 93.26 Mean value of accuracy, % 90.96 92.37 Speed (m/s) ϑ (t) Δ Journal of Robotics 7 Table 5: Comparative experimental results of diﬀerent detection methods on UAV aerial photography data. Method Ref.  Ref.  Ref.  Proposed method mAP, % 16.14 22.37 23.64 25.12 AP , % 26.60 34.29 38.08 39.75 AP , % 13.62 19.65 24.71 26.03 −1 Frame rate, frames·s 12 19 36 21 confusion. Its AP is 34.29%. Tao et al.  use rein- Conflicts of Interest forcement learning to search for drones, and the detection *e authors declare that there are no conﬂicts of interest process is easy to implement. *erefore, the detection speed −1 regarding the publication of this paper. is higher than the 21 frames·s of the proposed method. Its −1 frame rate is as high as 36 frames·s . However, due to the lack of PTZ camera assistance, it is not eﬀective in detecting References fast-moving targets. Its AP is 1.32% lower than the pro- posed method.  D. Chen, S. Sun, Z. Lei, H. Shao, and Y. Wang, “Ship target detection algorithm based on improved YOLOv3 for mari- time image,” Journal of Advanced Transportation, vol. 2021, 5. Conclusion p. 11, Article ID 9440212, 2021.  J. Abdulridha, Y. Ampatzidis, and S. C. Kakarla, “Detection of *e rapid development and application of UAV not only target spot and bacterial spot diseases in tomato using UAV- brings convenience to the society but also poses a serious based and benchtop-based hyperspectral imaging tech- threat to public security, personal privacy, and military niques,” Precision Agriculture, vol. 21, no. 5, pp. 1–24, 2020. security. It is becoming more and more important to ﬁnd  Z. Geng, R. Xu, and H. Deng, “LTE-based multistatic passive unknown UAVs quickly and accurately. *erefore, a radar system for UAV detection,” IET Radar, Sonar & target detection method of low-altitude UAV based on Navigation, vol. 14, no. 7, pp. 1088–1097, 2020. improved YOLOv3 network is proposed. *e improved  J. Yu, Y. Liu, Y. Bai, and F. Liu, “A double-threshold target YOLOv3 network is used to extract the boundary box of detection method in detecting low slow small target,” Procedia the target UAV. *e coordinate error and size error be- Computer Science, vol. 174, no. 5, pp. 616–624, 2020. tween the boundary box and the desired boundary box are  X. Chen, J. Guan, G. Wang, H. Ding, and Y. Huang, “Fast and reﬁned processing of radar maneuvering target based on calculated and input into the PID algorithm. *us, the hierarchical detection via sparse fractional representation,” center moving angle of the two-axis PTZ camera is IEEE Access, vol. 7, no. 7, pp. 149878–149889, 2019. controlled to realize the accurate detection of the target  T. Shi, H. Wang, W. Cui, and L. Ren, “Indoor space target UAV. *e experimental results based on the constructed searching based on EEG and EOG for UAV,” Soft Computing, dataset show that: vol. 23, no. 21, pp. 11199–11215, 2019.  L. Xiaoping, L. Songze, and Z. Boxing, “Fast aerial UAV (1) *e improved YOLOv3 network has better detection detection using improved inter-frame diﬀerence and SVM,” capabilities for UAV targets and is easy to distinguish Journal of Physics: Conference Series, vol. 1187, no. 2, similar objects. Compared with the YOLOv3 algo- pp. 032082–032093, 2019. rithm, the average detection accuracy rate is in-  R. *illainayagi and K. Senthil Kumar, “Combination of creased by about 1.5%. wavelet transform and singular value decomposition-based (2) *e proposed method uses the improved YOLOv3 contrast enhancement technique for target detection in UAV network to obtain the bounding box, and combines reconnaissance thermal images,” Journal of Modern Optics, vol. 66, no. 6, pp. 606–617, 2019. the PID algorithm to control the two-axis PTZ  Z. Li, C. Meng, F. Zhou et al., “Fast vision-based autonomous camera, which can better obtain the position of the detection of moving cooperative target for unmanned aerial drone. Its mAP, AP50, and AP75 are 25.12%, 39.75%, vehicle landing,” Journal of Field Robotics, vol. 36, no. 1, and 26.03%, respectively, and the frame rate is 21 −1 pp. 34–48, 2019. frames·s . *e overall detection performance is the  J. Abdulridha, Y. Ampatzidis, S. C. Kakarla, and P. Roberts, most ideal. “Detection of target spot and bacterial spot diseases in tomato using UAV-based and benchtop-based hyperspectral imaging In the next work, we will obtain more drone data sets in techniques,” Precision Agriculture, vol. 21, no. 5, pp. 955–978, scenarios for model training. A model with a faster speed and a smaller footprint will be selected for UAV identiﬁ-  F. Yang, B. Ma, J. Wang, H. Gao, and Z. Liu, “Target detection cation. It will allow the model to get rid of the dependence on of UAV aerial image based on rotational invariant depth the graphics card, making the cost of the system more denoising automatic encoder,” Xibei Gongye Daxue Xuebao/ aﬀordable. Journal of Northwestern Polytechnical University, vol. 38, no. 6, pp. 1345–1351, 2020. Data Availability  B. Jla, J. A. Rong, and L. A. Wei, “High precision detection algorithm based on improved RetinaNet for defect recogni- *e data used to support the ﬁndings of this study are in- tion of transmission lines - ScienceDirect,” Energy Reports, cluded within the article. vol. 6, no. 2, pp. 2430–2440, 2020. 8 Journal of Robotics  W. Xiaofei, “A grey wolf optimizer using Gaussian estimation of distribution and its application in the multi-UAV multi- target urban tracking problem - ScienceDirect,” Applied Soft Computing, vol. 78, no. 8, pp. 240–260, 2019.  X. Tao, S. Hou, and Y. Wang, “Feasibility of reinforcement learning for UAV-based target searching in a simulated communication denied environment,” Scientia Sinica Infor- mationis, vol. 50, no. 3, pp. 375–395, 2020.  X. Liu and Z. Zhang, “A vision-based target detection, tracking, and positioning algorithm for unmanned aerial vehicle,” Wireless Communications and Mobile Computing, vol. 2021, no. 7, pp. 1–12, 2021.  O. Doukhi, S. Hossain, and D.-J. Lee, “Real-time deep learning for moving target detection and tracking using unmanned aerial vehicle,” Journal of Institute of Control, Robotics and Systems, vol. 26, no. 5, pp. 295–301, 2020.  H.-J. Lim, S.-H. Choi, J. Oh, B. S. Kim, S. Kim, and J. H. Yang, “Adaptive ground control system of multiple-UAV operators in a simulated environment,” Aerospace Medicine and Human Performance, vol. 90, no. 10, pp. 841–850, 2019.  S. Perez-Carabaza ´ and J. Sch Erer, “UAV trajectory optimi- zation for Minimum Time Search with communication constraints and collision avoidance,” Engineering Applications of Artiﬁcial Intelligence, vol. 85, no. 11, pp. 357–371, 2019.  O. Kechagias-Stamatis and N. Aouf, “A new passive 3-D automatic target recognition architecture for aerial plat- forms,” IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 1, pp. 406–415, 2019.  D. Cavaliere, V. Loia, A. Saggese, S. Senatore, and M. Vento, “A human-like description of scene events for a proper UAV- based video content analysis,” Knowledge-Based Systems, vol. 178, no. 8, pp. 163–175, 2019.  A. A. d. Santos, J. Marcato Junior, M. S. Araujo ´ et al., “As- sessment of CNN-based methods for individual tree detection on images captured by RGB cameras attached to UAVs,” Sensors, vol. 19, no. 16, pp. 3595–3603, 2019.  M. Rosamilia, A. D. Maio, and A. Aubry, “Single-pulse si- multaneous target detection and angle estimation in a mul- tichannel phased array radar,” IEEE Transactions on Signal Processing, vol. 68, no. 7, pp. 6649–6664, 2020.  Z. Li, Q. Zhang, and T. Long, “Ship target detection and recognition method on sea surface based on multi-level hy- brid network,” Journal of Beijing Institute of Technology (Social Sciences Edition), vol. 30, no. 2, pp. 1–10, 2021.  E. Dong, B. Han, and H. Jian, “Moving target detection based on improved Gaussian mixture model considering camera motion,” Multimedia Tools and Applications, vol. 79, no. 11, pp. 7005–7020, 2020.  M. Li, G. Sun, J. Tong, and Z. He, “Training-free moving target detection with uncertain a priori knowledge for airborne radar,” IET Radar, Sonar & Navigation, vol. 14, no. 3, pp. 372–380, 2020.  A. Tz, A. Zp, and W. A. Hao, “Infrared small target detection via self-regularized weighted sparse model - ScienceDirect,” Neurocomputing, vol. 420, no. 3, pp. 124–148, 2021.
Journal of Robotics
Hindawi Publishing Corporation
Target Detection of Low-Altitude UAV Based on Improved YOLOv3 Network
Journal of Robotics
, Volume 2022 –
Mar 10, 2022
Share Full Text for Free
Add to Folder
Web of Science