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

Learn More →

A Novel Method to Reduce Time Investment When Processing Videos from Camera Trap Studies

A Novel Method to Reduce Time Investment When Processing Videos from Camera Trap Studies Camera traps have proven very useful in ecological, conservation and behavioral research. Camera traps non-invasively record presence and behavior of animals in their natural environment. Since the introduction of digital cameras, large amounts of data can be stored. Unfortunately, processing protocols did not evolve as fast as the technical capabilities of the cameras. We used camera traps to record videos of Eurasian beavers (Castor fiber). However, a large number of recordings did not contain the target species, but instead empty recordings or other species (together non-target recordings), making the removal of these recordings unacceptably time consuming. In this paper we propose a method to partially eliminate non-target recordings without having to watch the recordings, in order to reduce workload. Discrimination between recordings of target species and non-target recordings was based on detecting variation (changes in pixel values from frame to frame) in the recordings. Because of the size of the target species, we supposed that recordings with the target species contain on average much more movements than non-target recordings. Two different filter methods were tested and compared. We show that a partial discrimination can be made between target and non-target recordings based on variation in pixel values and that environmental conditions and filter methods influence the amount of non-target recordings that can be identified and discarded. By allowing a loss of 5% to 20% of recordings containing the target species, in ideal circumstances, 53% to 76% of non-target recordings can be identified and discarded. We conclude that adding an extra processing step in the camera trap protocol can result in large time savings. Since we are convinced that the use of camera traps will become increasingly important in the future, this filter method can benefit many researchers, using it in different contexts across the globe, on both videos and photographs. Citation: Swinnen KRR, Reijniers J, Breno M, Leirs H (2014) A Novel Method to Reduce Time Investment When Processing Videos from Camera Trap Studies. PLoS ONE 9(6): e98881. doi:10.1371/journal.pone.0098881 Editor: John Goodrich, Panthera, United States of America Received November 30, 2013; Accepted May 7, 2014; Published June 11, 2014 Copyright:  2014 Swinnen et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Funding: Research was funded by a Ph.D. grant of the Agency for Innovation by Science and Technology (IWT). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing Interests: The authors have declared that no competing interests exist. * E-mail: herwig.leirs@uantwerpen.be downloading of the images and the start of the statistical Introduction processing of data obtained from the images, is often overlooked. Camera traps, triggered by motion and/or heat of a passing The switch from analog to digital cameras and technical progress subject, are a non-invasive way to study animals and their such as increasing data storage capacity, caused the amount of behavior. Valuable knowledge is gathered by registering animals recordings to increase massively. As a consequence, processing in their natural habitat. Questions addressed by using camera recordings became more time consuming and is becoming one of traps are often related to animal ecology, behavior and conser- the limiting factors in the use of camera traps [29]. Automated vation [1]. For example, camera traps have been used to study image analysis has been used in different biological contexts niche separation [2], competitive exclusion [3], population ranging from microbiological ecology [30], experimental labora- structure [4,5], density estimation with [6,7] and without tory systems [31], phenotypic analysis [32], remote sensing [33] to individual recognition [8,9], abundance estimation [10], foraging corridor mapping [34]. However, publications about software behavior [11], biodiversity [12], activity patterns [13] and habitat solutions to manage and analyze image data and metadata use [14,15]. Camera traps can replace other study methods or add collected by camera traps are limited [29,35]. The TEAM to direct observations, track inventories, knowledge of local (Tropical Ecology Assessment & Monitoring) network uses a inhabitants or genetic surveys [16–20]. The target species are software package to manage and standardize image processing mostly medium to large terrestrial mammals since capture since they expect to have 1 000 000 photographs per year [36]. Yu probability decreases with decreasing size of the species [21,22]. et al. (2013) recently proposed a method to automatically recognize In recent years, species such as small arboreal primates [23,24], species, based on manually cropped camera trap photos of ectothermic Komodo dragons [25] and birds [26,27] have been animals, however, for their algorithm to work, they still had to subjects of camera trap study, showing wide ranging applicability. visually inspect and manually select all recordings [37]. The Rovero et al. (2013) showed that the amount of camera trap method proposed in this paper addresses the previous step in the papers being published in biological research is still increasing workflow, namely reducing the amount of recordings that must be [28]. Although a number of papers are published concerning visually inspected by automatically classifying the recordings methodology (see above), data processing, the step between the PLOS ONE | www.plosone.org 1 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos according to the potential of containing the target species (see nocturnal [41]. All movies were saved in the.AVI format on a Material and methods). Transcend SD HC 16GB card, and copied in the field to a small Although pictures are more commonly used in camera trap portable computer to a unique folder per camera and location. studies and are easier to process, videos provide more detailed Filename, date and hour of the recordings were automatically extracted to Excel. Location, camera number and species were information, especially behavioral. They are used in determining competitive exclusion [3], observing time budget [38], observing manually entered. An extra category was added cataloging image as ‘beaver’ or ‘non-target’ (including empty images and recordings behavior and determining population structure [4,39] and to study nest predation [40]. Videos are also more appealing to the general of other animals) since our main interest was to separate beaver recordings from all the rest. public when used as an awareness tool. Technological advances and continued innovation will ensure that camera traps will play Camera locations were divided according to the area of water visible in the video. The area was measured by using ImageJ, an an increasingly important role in wildlife studies [1]. We expect open source application to process images [42]. When the surface that the use of videos will increase and so will the need for a tool to automatically identify the non-target recordings. of the water was ,10% of the frame, videos were classified as Dry (5 locations, water surface ranges from 0–4%), while locations with In this study, we develop such a tool based on recordings .10% of water surface in the recordings were classified as Wet (5 gathered while studying beavers. We used camera traps to record locations, water surface ranges from 12%–57%). Two locations the spread of recently reintroduced beavers in northern Belgium, varied because of rainfall and drying of the water body respectively to evaluate territory occupancy by individuals or mated pairs and between 0% and 48% and 0% and 53%, and were included in the to study activity patterns. Wet classification since not enough beavers (n = 15) were Our first results indicate that, although the target species was registered to analyze these locations separately. recorded frequently, the majority of the recordings were empty or We were ultimately interested in recordings of beavers. But contained non-target species. Since it is very time consuming to using computer algorithms to discriminate beavers in video watch all recordings, we developed an automated filtering method. footage was a very difficult endeavor, requiring high-level pattern The goal of this algorithm is to eliminate the maximum amount of recognition, which was by no means the aim of this study. Our the non-target recordings while minimizing the amount of target goal was less ambitious, as we tried to discriminate the videos that recordings discarded. To our knowledge, this extra filtering of were likely to be beaver free (non-target recordings) and to reduce recordings (photos or videos) between the downloading of the the set of videos that had to be inspected for beaver presence. As a recordings and the manually processing was never reported before consequence, we tried to maximize the true positive rate (TP-rate; in camera trap studies. number of non-target recordings correctly classified as non-target recordings divided by total number of non-target recordings) while Materials and Methods minimizing the false positive rate (FP-rate; number of target Bushnell Trophy cams (Bushnell Outdoor Products, 9200 Cody, recordings wrongly classified as non-target recordings divided by Overland Park, Kansas 66214, Model 119436c) were positioned at total number of target recordings) (Table 1). 12 different locations in 9 different beaver territories, in the The discrimination was based on to what extent the video province of Limburg, in the east of Flanders, 20 July 2012 - 8 frames change throughout the length of the video. As beavers are October 2012. The responsible authority, the ‘Agentschap voor fairly large mammals (0.80–1.20 m body length, 0.25–0.50 m tail Natuur en Bos’ (Agency for Nature and Forest) decided that, length, 11–30 kg [43]) and among the largest in our study area, it although beavers are a protected species, no special permit was is to be expected that their presence on the footage will induce required since camera traps do not disturb the animals. bigger changes, compared to other smaller animals, e.g. small Permissions to access the camera trap locations were granted by rodents and birds, or movement of water and/or vegetation Limburgs Landschap (1 location), Natuurpunt (5 locations), nv De registered in the recordings. In the following, we propose and Scheepvaart (1 location) and Steengoed (5 locations). evaluate two different ways to quantify the amount of ‘move- Cameras were attached to a nearby tree 30–90 cm above the ments’, on which the discrimination will be based. All following ground and directed at the burrow (5 locations), a running track (3 manipulations and calculations were performed in MATLAB [44] locations), or a foraging location (4 locations). The anticipated except when stated differently. passage of a beaver was never farther than 6 m away from the To start, we performed two basic manipulations of the camera (but often closer). The camera settings were standardized recordings. The first two frames were removed from each movie over all cameras as follows: Type = Video, Video Size = 7206480, because of the time stamp on the first frame and the instability of Video Length = 15s, Interval = 1s, Time Stamp = On, Video the light caused by the starting of the camera in the first two sound = On. Cameras were activated when detecting a combina- frames. Second, we averaged out small spatial and temporal tion of movement and heat. Cameras were positioned for an changes due to detection noise, movement of vegetation, water average of 35 days at each location (range = 30 to 50 days, reflections, etc. This was done by down sampling (averaging) the SD = 5.6). The sensitivity of the sensor was set to low, medium or video along both the spatial, x and y, and temporal dimension, t high, according to local circumstances. The medium sensitivity (respectively with factor 5, 5, 10). As beavers are large and move was used in most environmental settings. When cameras were rather slowly, the detection of their movement would not suffer directed to highly dynamic streams, the sensitivity was set low. from this reduction. Moreover, this operation would also improve When beavers were expected to pass rather far from the camera computational speed. and vegetation was limited, high sensitivity was selected. The To detect non-target recordings, we first quantified the amount sensitivity must be sufficiently high since we suspect that the cold of variation throughout the video. If the pixel values did not vary water in the fur of the beavers reduces the probability of detection much throughout the video, then it was unlikely that the camera by camera traps. The illumination of recordings in poor light was triggered by an animal as large as a beaver. We used two conditions (dusk, dawn or night) was assisted by infrared LEDs, different methods to quantify this amount of pixel variation and resulting in black and white recordings. Only black and white arrived at a ‘measure of frame variation’. Both were chosen since recordings were considered since it is known that beavers are they are easy to apply and differ qualitatively from each other. PLOS ONE | www.plosone.org 2 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos Table 1. Classification matrix of the recordings. Reality Non-target recordings Target recordings Classification result Non-Target recordings True Positive False Positive Target recordings False Negative True Negative 1991 videos were recorded at 12 different locations in 9 different beaver territories, in the province of Limburg, in the east of Flanders, Belgium, between 20 July 2012 and 8 October 2012. We recorded 1043 recordings of the target species, the beaver, 553 empty recordings and 395 recordings of non-target species. Every recording was classified based on D = ‘‘the amount of pixel variation’’ as target or non-target recording. The correct classification of non-target recordings was considered to be a success (True positive, TP) since these recordings can be correctly discarded. This value must be as high as possible in order to remove the maximum amount of non- target recordings. False positives (FP) were the beavers (target recordings) which were classified as non-target recordings and wrongly discarded. This number must be as small as possible since valuable data is being discarded. False negatives (FN) were non-target recordings which were classified as being target recordings. True negatives (TN) were the target recordings which were recognized as being target recordings. doi:10.1371/journal.pone.0098881.t001 In the first method, Filter 1, we considered the frame variations the total number of non-target recordings) and the false positive rate (FP-rate; the number of beaver recordings misclassified as compared to the average frame V(x,y)~SV(x,y; t)T , of which non-target, divided by the total number of beaver recordings) that every pixel (x,y) was averaged across the length of the video (along a classification based on this threshold would produce. Variation t = time). At every time step the squared distance of the frame of the threshold over the full range of D-values, resulted in the so- V(x,y; t) to the average frame is calculated, called receiver operating characteristics curve (ROC curve, [45]), from which the costs and benefits corresponding to a particular d (t)~ V(x,y; t){V(x,y) threshold could be easily assessed. This analysis was performed in x,y RStudio [46]. We also calculated the time savings generated by using a filter We used the following equation as a measure of the ‘amount of method, as the summed duration of the recordings that would not pixel variation’: have to be inspected visually. To determine the ability of the threshold to discard the intended amount of non-target videos in 2 2 spite of sampling variation (i.e. proportion of beavers discarded in D~ max d (t) { min d (t) a a future videos), a bootstrap procedure was carried out. Briefly, a subset of 500 observations were resampled from the dataset and classified according to the original 5% FP-rate. The operation was Note that for the calculation of D we subtracted min d (t) in repeated 1000 times and the FP-ratio was calculated at each step. order to compensate for the fact that different environments show Descriptive statics for the bootstrapped FP-ratio were computed different base line variations, e.g. aquatic sceneries exhibit more for the Complete, Dry and Wet datasets. variations compared to dry area and we were interested in the movements of animals against this variable background. Results The second method, Filter 2, focused on the changes between subsequent frames During 405 camera nights, the 12 cameras recorded 1991 videos, 933 recordings in dry locations and 1058 in wet locations, 2 2 d (t)~ ½ V(x,y; t){V(x,y; t{1) , with a mean of 166 recordings per location (49–296, SD = 81.9). x,y All recordings were watched and cameras registered 1043 recordings of the target species, the beaver, 553 empty recordings and is essentially different from the previous method. The and 395 recordings of non-target species. The resulting ROC- parameter D was calculated in the same way as before (see Matlab Code S1). curves are shown in Figure 1 for the two different filters, when applying them to videos recorded in different conditions. Both Both methods quantify the changes in pixel values within the filter methods discriminated between non-target and potential recording, which we consider to be a proxy for movement. If this D value is fairly low, then the recorded movements were rather target recordings, but Filter 2 performed slightly better than Filter 1, especially when lower FP-rates were tolerated (5% FP-rate small; a larger value points to increased activity during the 15 marked by a vertical dashed line). Also, irrespective of the FP-rate, second recording. These calculated values were used to discrim- inate the assumed non-target recordings from the possible target the TP-rate in Dry circumstances was always higher compared to Wet circumstances. recordings. The FP-rate that can be tolerated depends on the study goals This discrimination was done by means of a threshold: if the calculated D-value of the video was below this threshold, then we and design, and will be most decisive in what threshold can be used. In this particular study, a loss of 5% to 20% of the beaver assumed that the video was empty (a non-target recording); we did not make any inferences on the target’s presence when the footage was tolerable. A FP-rate of 5% resulted in a reduction of threshold value was exceeded. Hence, our approach was aimed at workload allowing us to remove 26% to 53% of non-target recordings (time savings 30–53 min) in dry conditions, 13% to detecting non-target recordings. Classifying a video without a beaver (empty or with other species) as a non-target video was 33% (time savings 25–55 min) in wet conditions and 18% to 42% considered a success, see Table 1. Consequently, for a particular (time savings 56–113 min) in the complete dataset. When we threshold value, one could calculate the true positive rate (TP-rate; tolerated a FP-rate of 20%, then 72% to 76% (time savings 92– the number of correctly classified non-target recordings divided by 95 min) of non-target recordings could be discarded in dry PLOS ONE | www.plosone.org 3 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos Figure 1. Possible gain (true positive rate, TP-rate) given an accepted loss (false positive rate, FP-rate). The FP-rate represents the proportion of target recordings (beavers) classified as non-target recordings. The TP-rate is the proportion of non-target recordings correctly classified as non-target. This is the proportion of non-target recordings that will be discarded correctly given a certain FP-rate. The best performing filter maximizes the TP-rate while minimizing the FP- rate. Filter 2 performs better in all environmental circumstances. The dashed diagonal represents the outcome of a random model which cannot discriminate between target and non-target recordings. The dashed vertical line represents a 5% threshold (FP-rate). Dry,10% water area in footage (5 locations, n = 933 recordings), Wet.10% water area (7 locations, n = 1058 recordings), Complete dataset is the combined Dry and Wet dataset (12 locations, n = 1991 recordings). doi:10.1371/journal.pone.0098881.g001 conditions, but only 46% to 54% (time savings 92–104 min) in wet conditions and between 59% and 65% (time savings 191–206 min) in the complete dataset. The bootstrap analysis indicated that sampling variation had a limited effect on the FP-ratio (Table 2). Within the 95% confidence interval, a minimum of 2.9% and a maximum of 7.4% of beaver images would be discarded using the 5% threshold, based on the complete dataset. Results of the bootstrap analysis indicated that sampling variation had only a limited effect in all environmental circumstances on the percentage of record- ings discarded. Discussion We show that a filter method based on changing pixel values can partly discriminate between recordings of the study species and non-target recordings. The amount of non-target recordings that can be discarded without watching the footage depends on the chosen threshold and this threshold will vary between studies. In both filter methods, it is clear that results depend on the environmental circumstances of the camera locations. It is easier to distinguish between beaver and non-target recordings in Dry circumstances because variation is lower than in Wet circum- stances. The filtering mechanism will work best on medium to large mammals, but these are also the most suitable subjects of species inventory studies by camera traps [47]. Since these filter PLOS ONE | www.plosone.org 4 June 2014 | Volume 9 | Issue 6 | e98881 Table 2. A bootstrap analysis was performed for both filter methods (Filter 1 and Filter 2) on the complete dataset (n = 1991), on the videos recorded at dry locations (n = 933) and on the videos recorded at wet locations (n = 1058). Dry Wet Complete dataset Filter 1 Filter 2 Filter 1 Filter 2 Filter 1 Filter 2 False Positive Rate Mean 0.051 0.051 0.050 0.051 0.051 0.051 Sd 0.009 0.009 0.011 0.010 0.012 0.011 2.5 percentile 0.035 0.035 0.029 0.032 0.029 0.029 97.5 percentile 0.068 0.067 0.072 0.072 0.074 0.074 A subsample of 500 videos was randomly sampled with replacement and this was repeated 1000 times. Recordings were classified based on their D-value; the thresholds were chosen to result in a 5% false positive rate in the respective datasets (see text). The new mean, standard deviation (Sd), 2.5 and 97.5 percentile are shown. doi:10.1371/journal.pone.0098881.t002 Novel Filter Mechanism for Camera Trap Videos methods are the first attempt to partially automatically process [4,48,49] it is probably not acceptable to miss a recording since a recordings, we did not aim to create a very complicated high level single recording may have a large impact on results. For this study, pattern recognition program. The current method is rather robust a loss of 5% to 20% of the beaver footage was tolerable, since to changes in camera-angle, illumination, colour or black and beavers were recorded frequently and the data were collected to white recordings, distance from the camera to the subject, size of determine average activity patterns. A reduction of the tolerated the subject and the position of the subject to the camera since they loss of target recordings would have lowered the TP-rate, resulting can be largely accounted for by choice of the threshold. in less time savings. How quickly the TP-rate decreases with The greatest differences between filter methods were observed decreasing FP-rate depends on the shape of the ROC-curve and when the tolerated FP-rate was small. When using a FP-rate of will differ between studies. Also, these filter methods are not 5%, Filter 2 resulted in a TP–rate which was more than double the suitable for species which are very hard to record because it will value of Filter 1, in all environmental circumstances. When take a very long time before enough data are collected to increasing the tolerated FP-rate, these differences became smaller determine the threshold. (Figure 1). The filter protocol is optimized for processing videos and not Discarding up to 76% of non-target images can prove to be very photographs. The same process could be used for processing timesaving when used in long term survey studies deploying large photographs. Although each picture is a still image, the number of cameras. This reduction of workload has two direct comparison of consecutive pictures (as if they were frames of a implications for camera trap studies. Since the processing of movie) makes the analysis possible. The disadvantage of this recordings takes less time, the number of cameras can be increased method is that the time between pictures can vary a lot when allowing covering a larger study area or augmenting the amount of cameras are movement-triggered, and environmental factors, like cameras used in the same study area so that patterns can be growth of plants or changing light conditions, can make it more discerned on a finer scale. Second, the sensitivity of the camera difficult to compare different photographs. However, Hamel et al. traps can be increased, allowing the cameras to react to smaller (2013) showed that using a fixed 5-min interval resulted in lower animals. This will result in an increase of the number of daily presence raw error rates compared to movement-triggered recordings, but the time spent on processing the images will still cameras and recommend opting for time-triggered cameras when be reduced because of the filtering of non-target recordings. The aiming to capture abundant species [50]. When using fixed time recordings of small (non-target) animals may be useful for other intervals, high variation is avoided, making it easier to process studies. The same database of video recordings could then be used, images when the interval between pictures is not too long. but now with the small animal as target, without having to perform However, problems can still occur when light conditions change a new field campaign. rapidly as during sunset and sunrise. When using a series of Although we show that the proposed filter method can reduce pictures that are taken with fast intervals, as a consequence of a the amount of work substantially given a certain cost (lost of single trigger event, this problem can be avoided. recordings of interest), there are limitations to this method. Camera traps are imperfect detectors and the chance of First, a number of recordings must be viewed to determine if the detecting a species decreases with decreasing species size and species of interest is recorded. Only when a sufficient number of increasing speed and is influenced by seasons [22,51]. But a recordings of the species of interest is obtained, a comparison camera can also be ‘tricked’ into recording when there are no between these and other images can be made. The necessary animals. This generates empty recordings and these are rarely amount of recordings depends highly on the studied species (size reported although they give an important indication about the and speed), other sympatric species and the environmental efficiency and expected time effort necessary to process all conditions. Larger species that are detected regularly and will be recordings. The reasons for recording empty recordings are studied for a long time by using camera traps are the most suitable diverse, moving vegetation, technical problems, moving water, subjects to discriminate from non-target recordings. Based on the sunlight reflection or time lag between passing of an animal and determined threshold, images which most likely not contain starting of the camera. These factors will vary depending on animals can be filtered out, but a distinction between different environment, season and study species. These problems are well species of similar size and speed is not feasible. known to researchers but receive only limited attention in the The filter methods are based on changing pixel values. Animals literature [52]. To reduce the amount of empty recordings, the that do not move (stay on the same location during the length of sensitivity of the sensor can be reduced. However, this also the complete recording) will most likely be discarded. It is however increases the chance of missing the study species. Although no reasonable to assume that a motionless animal was detected while published research is available, we think that sensor sensitivity arriving or will be detected while leaving the location where the must be augmented in semi-aquatic mammals when being filmed camera trap is aimed at. Animals must be of sufficient size to stand on land since they are likely to just have left the water, resulting in out from the background noise and to be recognized as a potential a cold layer around the body of the animal making it more difficult target recording. For very small species, it may be necessary to skip to detect than a similar sized dry mammal. This high sensitivity the down sampling step in both spatial directions in order to results in more false detections, explaining the high percentage of achieve an image that is detailed enough. For very fast species, the empty recordings compared to other studies 15,7% [20]. We want down sampling in the temporal dimension can be detrimental. to encourage authors to not only mention the trapping effort and However, the performance of the method depends highly on the the amount of pictures of the studied species but also the amount variation (movement) in the background. Once the variation in the of empty recordings (or recordings of non-target species) making it background is larger than the changes induced by passing species, possible to compare how different types of cameras perform in it will not be possible to distinguish between target and non-target different circumstances allowing future researchers to benefit from recordings. this knowledge. The use of the filter protocol must be considered with Although we acknowledge the limitations of our filter method, knowledge of the species, species community and the study design. we believe an important progress has been made by showing that For example: when camera traps are used to perform capture adding an extra filter step between downloading of images and the mark recapture analysis on individually recognizable individuals PLOS ONE | www.plosone.org 5 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos locations. We thank Jemp Peeters, Nol Goossens and Roger Nijssen for statistical processing can save lots of valuable time, with losing only providing valuable help in fieldwork. We thank 1 anonymous referee and a limited amount of data. John Goodrich for the helpful comments and suggestions that considerably improved the manuscript. Supporting Information Code S1 Matlab Code. Author Contributions (M) Conceived and designed the experiments: KRRS JR MB HL. Performed the experiments: KRRS JR MB. Analyzed the data: KRRS JR MB. Wrote Acknowledgments the paper: KRRS JR MB HL. We would like to thank Limburgs Landschap, Natuurpunt, nv De Scheepvaart and Steengoed for granting permission to access camera trap References 1. O’Connell AF, Nichols JD, Karanth UK (2011) Camera Traps in Animal 25. Ariefiandy A, Purwandana D, Seno A, Ciofi C, Jessop TS (2013) Can Camera Ecology Methods and Analyses. New York: Springer. Traps Monitor Komodo Dragons a Large Ectothermic Predator? PLoS One 8: 2. De Almeida Jacomo AT, Silveira L, Diniz-Filho JAF (2004) Niche separation 1–8. between the maned wolf (Chrysocon brachyurus), the crab-eating fox (Dusicyon thous) 26. O’Brien TG, Kinnaird MF (2008) A picture is worth a thousand words: the and the hoary fox (Dusicyon vetulus) in central Brazil. J Zool 262: 99–106. application of camera trapping to the study of birds. Bird Conserv Int 18: 144– 3. Head JS, Robbins MM, Mundry R, Makaga L, Boesch C (2012) Remote video- 162. camera traps measure habitat use and competitive exclusion among sympatric 27. Seki S-I (2010) Camera-trapping at artificial bathing sites provides a snapshot of chimpanzee, gorilla and elephant in Loango National Park, Gabon. J Trop Ecol a forest bird community. J For Res 15: 307–315. 28: 571–583. 28. Rovero F, Zimmermann F, Berzi D, Meek P (2013) ‘‘ Which camera trap type 4. Hariyadi ARS, Priambudi A, Setiawan R, Daryan D, Yayus A, et al. (2011) and how many do I need?’’ A review of camera features and study designs for a Estimating the population structure of Javan rhinos (Rhinoceros sondaicus) in Ujung range of wildlife research applications. Hystrix, Ital J Mammology 24: 148-156. Kulon National Park using the mark- recapture method based on video and 29. Harris G, Thomson R, Childs JL, Sanderson JG (2010) Automatic storage and camera trap identification. Pachyderm: 90–99. analysis of camera trap data. Bull Ecol Soc Am: 352–360. 5. Karanth KU, Nichols JD, Kumar NS, Hines JE (2006) Assessing tiger 30. Daims H, Lu ¨ cker S, Wagner M (2006) Daime, a Novel Image Analysis Program population dynamics using photographic capture-recapture sampling. Ecology for Microbial Ecology and Biofilm Research. Environ Microbiol 8: 200–213. 87: 2925–2937. 31. Pennekamp F, Schtickzelle N (2013) Implementing image analysis in laboratory- 6. Royle AJ, Nichols JD, Karanth UK, Gopalaswamy AM (2009) A hierarchical based experimental systems for ecology and evolution: a hands-on guide. model for estimating density in camera-trap studies. J Appl Ecol 46: 118–127. Methods Ecol Evol 4: 483–492. 7. Karanth UK, Nichols JD (1998) Estimation of tiger densities in India using 32. Vogt A, Cholewinski A, Shen X, Nelson SG, Lazo JS, et al. (2009) Automated photographic captures and recaptures. Ecology 79: 2852–2862. image-based phenotypic analysis in zebrafish embryos. Dev Dyn 238: 656–663. 8. Rovero F, Marshall AR (2009) Camera trapping photographic rate as an index 33. Addink EA., De Jong SM, Davis SA., Dubyanskiy V, Burdelov LA., et al. (2010) of density in forest ungulates. J Appl Ecol 46: 1011–1017. The use of high-resolution remote sensing for plague surveillance in Kazakhstan. 9. Rowcliffe MJ, Field J, Turvey ST, Carbone C (2008) Estimating animal density Remote Sens Environ 114: 674–681. using camera traps without the need for individual recognition. J Appl Ecol 45: 34. Vogt P, Riitters KH, Iwanowski M, Estreguil C, Kozak J, et al. (2007) Mapping 1228–1236. landscape corridors. Ecol Indic 7: 481–488. 10. Marnewick K, Funston PJ, Karanth KU (2008) Evaluating camera trapping as a 35. Sundaresan SR, Riginos C, Abelson ES (2011) Management and Analysis of method for estimating cheetah abundance in ranching areas. South Camera Trap Data; Alternative Approaches (Response to Harris et al., 2010). African J Wildl Res 38: 59–65. Bull Ecol Soc Am: 188–195. 11. Prasad S, Pittet A, Sukumar R (2009) Who really ate the fruit? A novel approach 36. Fegraus EH, Lin K, Ahumada JA, Baru C, Chandra S, et al. (2011) Data to camera trapping for quantifying frugivory by ruminants. Ecol Res 25: 225– acquisition and management software for camera trap data: A case study from the TEAM Network. Ecol Inform 6: 345–353. 12. O’Brien TG, Baillie JEM, Krueger L, Cuke M (2010) The Wildlife Picture 37. Yu X, Wang J, Kays R, Jansen P a, Wang T, et al. (2013) Automated Index: monitoring top trophic levels. Anim Conserv 13: 335–343. identification of animal species in camera trap images. EURASIP J Image Video 13. Bridges AS, Vaughan MR, Klenzendorf S (2004) Seasonal variation in Process 52. American black bear Ursus americanus activity patterns: quantification via remote 38. Scheibe KM, Eichhorn K, Wiesmayr M, Schonert B, Krone O (2007) Long- photography. Wildlife Biol 4: 277–284. term automatic video recording as a tool for analysing the time patterns of 14. Di Bitetti MS, Paviolo A, De Angelo C (2006) Density, habitat use and activity utilisation of predefined locations by wild animals. Eur J Wildl Res 54: 53–59. patterns of ocelots (Leopardus pardalis) in the Atlantic Forest of Misiones, 39. Hariyadi AR, Setiawan R, Yayus DA, Purnama H (2010) Preliminary behaviour Argentina. J Zool 270: 153–163. observations of the Javan rhinoceros (Rhinoceros sondaicus) based on video trap 15. Tobler MW, Carrillo-Percastegui SE, Powell G (2009) Habitat use, activity surveys in Ujung Kulon National Park. Pachyderm: 93–99. patterns and use of mineral licks by five species of ungulate in south-eastern 40. Staller EL, Palmer WE, Carroll JP, Thornton RP, Sisson DC (2005) Identifying Peru. J Trop Ecol 25: 261. Predators At Northern Bobwhite Nests. J Wildl Manage 69: 124–132. 16. Janec ˇka JE, Munkhtsog B, Jackson RM, Naranbaatar G, Mallon DP, et al. 41. Sharpe F, Rosell F (2003) Time budgets and sex differences in the Eurasian (2011) Comparison of noninvasive genetic and camera-trapping techniques for beaver. Anim Behav 66: 1059–1067. surveying snow leopards. J Mammal 92: 771–783. 42. Schneider C a, Rasband WS, Eliceiri KW (2012) NIH Image to ImageJ: 25 17. Srbek-Araujo AC, Chiarello AG (2005) Is camera-trapping an efficient method years of image analysis. Nat Methods 9: 671–675. for surveying mammals in Neotropical forests? A case study in south-eastern 43. Macdonald DW (2001) Beavers. The new encyclopedia of mammals. Oxford: Brazil. J Trop Ecol 21: 121–125. Oxford University Press. pp. 590–595. 18. O’Brien TG, Kinnaird MF, Wibisono HT (2003) Crouching tigers, hidden prey: 44. Matlab (2012). Version R2012b. 8.0.0.783. The MathWorks Inc. Natick Sumatran tiger and prey populations in a tropical forest landscape. Anim Massachusetts. Conserv 6: 131–139. 45. Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett 27: 19. Silveira L, Ja ´ como AT a., Diniz-Filho JAF (2003) Camera trap, line transect 861–874. census and track surveys: a comparative evaluation. Biol Conserv 114: 351–355. ¨ 46. Rstudio (2012). RStudio: Integrated development environment for R (Version 20. Can OE, Togan I (2009) Camera trapping of large mammals in Yenice Forest 0.96.122) [Computer software]. Boston, MA. Available http://www.rstudio. Turkey local information versus camera traps. Oryx 43: 427–430. org/. Accessed 2014 Jan 10. 21. Tobler MW, Carrillo-Percastegui SE, Leite Pitman R, Mares R, Powell G (2008) 47. Kelly MJ (2008) Design, evaluate, refine: camera trap studies for elusive species. Further notes on the analysis of mammal inventory data collected with camera Anim Conserv 11: 182–184. traps. Anim Conserv 11: 187–189. 48. Soisalo MK, Cavalcanti SMC (2006) Estimating the density of a jaguar 22. Rowcliffe MJ, Carbone C, Jansen PA, Kays R, Kranstauber B (2011) population in the Brazilian Pantanal using camera-traps and capture–recapture Quantifying the sensitivity of camera traps: an adapted distance sampling sampling in combination with GPS radio-telemetry. Biol Conserv 129: 487–496. approach. Methods Ecol Evol 2: 464–476. 23. Olson ER, Marsh RA, Bovard BN, Randrianarimanana LHL, Ravaloharima- 49. Carbone C, Christie S, Conforti K, Coulson T, Franklin N, et al. (2001) The use nitra M, et al. (2012) Arboreal camera trapping for the Critically endangered of photographic rates to estimate densities of tigers and other cryptic mammals. greater bamboo lemur Prolemur simus. Oryx 46: 593–597. Anim Conserv 4: 75–79. 24. Tan CL, Yang Y, Niu K (2013) Into the night: camera traps reveal nocturnal 50. Hamel S, Killengreen ST, Henden J-A, Eide NE, Roed-Eriksen L, et al. (2013) activity in a presumptive diurnal primate, Rhinopithecus brelichi. Primates 54: 1–6. Towards good practice guidance in using camera-traps in ecology: influence of PLOS ONE | www.plosone.org 6 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos sampling design on validity of ecological inferences. Methods Ecol Evol 4: 105– 52. Swann DE, Kawanishi K, Palmer J (2011) Evaluating types and features of camera traps in ecological studies: a guide for researchers. In: O’Connell AF, 51. Tobler MW, Carrillo-Percastegui SE, Leite Pitman R, Mares R, Powell G (2008) Nichols JD, Karanth KU, editors. Camera Traps in Animal Ecology:Methods An evaluation of camera traps for inventorying large- and medium-sized and Analyses. New York: Springer. pp. 27–43. terrestrial rainforest mammals. Anim Conserv 11: 169–178. PLOS ONE | www.plosone.org 7 June 2014 | Volume 9 | Issue 6 | e98881 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png PLoS ONE Pubmed Central

A Novel Method to Reduce Time Investment When Processing Videos from Camera Trap Studies

PLoS ONE , Volume 9 (6) – Jun 11, 2014

Loading next page...
 
/lp/pubmed-central/a-novel-method-to-reduce-time-investment-when-processing-videos-from-legUZTorja

References (57)

Publisher
Pubmed Central
Copyright
© 2014 Swinnen et al
ISSN
1932-6203
eISSN
1932-6203
DOI
10.1371/journal.pone.0098881
Publisher site
See Article on Publisher Site

Abstract

Camera traps have proven very useful in ecological, conservation and behavioral research. Camera traps non-invasively record presence and behavior of animals in their natural environment. Since the introduction of digital cameras, large amounts of data can be stored. Unfortunately, processing protocols did not evolve as fast as the technical capabilities of the cameras. We used camera traps to record videos of Eurasian beavers (Castor fiber). However, a large number of recordings did not contain the target species, but instead empty recordings or other species (together non-target recordings), making the removal of these recordings unacceptably time consuming. In this paper we propose a method to partially eliminate non-target recordings without having to watch the recordings, in order to reduce workload. Discrimination between recordings of target species and non-target recordings was based on detecting variation (changes in pixel values from frame to frame) in the recordings. Because of the size of the target species, we supposed that recordings with the target species contain on average much more movements than non-target recordings. Two different filter methods were tested and compared. We show that a partial discrimination can be made between target and non-target recordings based on variation in pixel values and that environmental conditions and filter methods influence the amount of non-target recordings that can be identified and discarded. By allowing a loss of 5% to 20% of recordings containing the target species, in ideal circumstances, 53% to 76% of non-target recordings can be identified and discarded. We conclude that adding an extra processing step in the camera trap protocol can result in large time savings. Since we are convinced that the use of camera traps will become increasingly important in the future, this filter method can benefit many researchers, using it in different contexts across the globe, on both videos and photographs. Citation: Swinnen KRR, Reijniers J, Breno M, Leirs H (2014) A Novel Method to Reduce Time Investment When Processing Videos from Camera Trap Studies. PLoS ONE 9(6): e98881. doi:10.1371/journal.pone.0098881 Editor: John Goodrich, Panthera, United States of America Received November 30, 2013; Accepted May 7, 2014; Published June 11, 2014 Copyright:  2014 Swinnen et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Funding: Research was funded by a Ph.D. grant of the Agency for Innovation by Science and Technology (IWT). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing Interests: The authors have declared that no competing interests exist. * E-mail: herwig.leirs@uantwerpen.be downloading of the images and the start of the statistical Introduction processing of data obtained from the images, is often overlooked. Camera traps, triggered by motion and/or heat of a passing The switch from analog to digital cameras and technical progress subject, are a non-invasive way to study animals and their such as increasing data storage capacity, caused the amount of behavior. Valuable knowledge is gathered by registering animals recordings to increase massively. As a consequence, processing in their natural habitat. Questions addressed by using camera recordings became more time consuming and is becoming one of traps are often related to animal ecology, behavior and conser- the limiting factors in the use of camera traps [29]. Automated vation [1]. For example, camera traps have been used to study image analysis has been used in different biological contexts niche separation [2], competitive exclusion [3], population ranging from microbiological ecology [30], experimental labora- structure [4,5], density estimation with [6,7] and without tory systems [31], phenotypic analysis [32], remote sensing [33] to individual recognition [8,9], abundance estimation [10], foraging corridor mapping [34]. However, publications about software behavior [11], biodiversity [12], activity patterns [13] and habitat solutions to manage and analyze image data and metadata use [14,15]. Camera traps can replace other study methods or add collected by camera traps are limited [29,35]. The TEAM to direct observations, track inventories, knowledge of local (Tropical Ecology Assessment & Monitoring) network uses a inhabitants or genetic surveys [16–20]. The target species are software package to manage and standardize image processing mostly medium to large terrestrial mammals since capture since they expect to have 1 000 000 photographs per year [36]. Yu probability decreases with decreasing size of the species [21,22]. et al. (2013) recently proposed a method to automatically recognize In recent years, species such as small arboreal primates [23,24], species, based on manually cropped camera trap photos of ectothermic Komodo dragons [25] and birds [26,27] have been animals, however, for their algorithm to work, they still had to subjects of camera trap study, showing wide ranging applicability. visually inspect and manually select all recordings [37]. The Rovero et al. (2013) showed that the amount of camera trap method proposed in this paper addresses the previous step in the papers being published in biological research is still increasing workflow, namely reducing the amount of recordings that must be [28]. Although a number of papers are published concerning visually inspected by automatically classifying the recordings methodology (see above), data processing, the step between the PLOS ONE | www.plosone.org 1 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos according to the potential of containing the target species (see nocturnal [41]. All movies were saved in the.AVI format on a Material and methods). Transcend SD HC 16GB card, and copied in the field to a small Although pictures are more commonly used in camera trap portable computer to a unique folder per camera and location. studies and are easier to process, videos provide more detailed Filename, date and hour of the recordings were automatically extracted to Excel. Location, camera number and species were information, especially behavioral. They are used in determining competitive exclusion [3], observing time budget [38], observing manually entered. An extra category was added cataloging image as ‘beaver’ or ‘non-target’ (including empty images and recordings behavior and determining population structure [4,39] and to study nest predation [40]. Videos are also more appealing to the general of other animals) since our main interest was to separate beaver recordings from all the rest. public when used as an awareness tool. Technological advances and continued innovation will ensure that camera traps will play Camera locations were divided according to the area of water visible in the video. The area was measured by using ImageJ, an an increasingly important role in wildlife studies [1]. We expect open source application to process images [42]. When the surface that the use of videos will increase and so will the need for a tool to automatically identify the non-target recordings. of the water was ,10% of the frame, videos were classified as Dry (5 locations, water surface ranges from 0–4%), while locations with In this study, we develop such a tool based on recordings .10% of water surface in the recordings were classified as Wet (5 gathered while studying beavers. We used camera traps to record locations, water surface ranges from 12%–57%). Two locations the spread of recently reintroduced beavers in northern Belgium, varied because of rainfall and drying of the water body respectively to evaluate territory occupancy by individuals or mated pairs and between 0% and 48% and 0% and 53%, and were included in the to study activity patterns. Wet classification since not enough beavers (n = 15) were Our first results indicate that, although the target species was registered to analyze these locations separately. recorded frequently, the majority of the recordings were empty or We were ultimately interested in recordings of beavers. But contained non-target species. Since it is very time consuming to using computer algorithms to discriminate beavers in video watch all recordings, we developed an automated filtering method. footage was a very difficult endeavor, requiring high-level pattern The goal of this algorithm is to eliminate the maximum amount of recognition, which was by no means the aim of this study. Our the non-target recordings while minimizing the amount of target goal was less ambitious, as we tried to discriminate the videos that recordings discarded. To our knowledge, this extra filtering of were likely to be beaver free (non-target recordings) and to reduce recordings (photos or videos) between the downloading of the the set of videos that had to be inspected for beaver presence. As a recordings and the manually processing was never reported before consequence, we tried to maximize the true positive rate (TP-rate; in camera trap studies. number of non-target recordings correctly classified as non-target recordings divided by total number of non-target recordings) while Materials and Methods minimizing the false positive rate (FP-rate; number of target Bushnell Trophy cams (Bushnell Outdoor Products, 9200 Cody, recordings wrongly classified as non-target recordings divided by Overland Park, Kansas 66214, Model 119436c) were positioned at total number of target recordings) (Table 1). 12 different locations in 9 different beaver territories, in the The discrimination was based on to what extent the video province of Limburg, in the east of Flanders, 20 July 2012 - 8 frames change throughout the length of the video. As beavers are October 2012. The responsible authority, the ‘Agentschap voor fairly large mammals (0.80–1.20 m body length, 0.25–0.50 m tail Natuur en Bos’ (Agency for Nature and Forest) decided that, length, 11–30 kg [43]) and among the largest in our study area, it although beavers are a protected species, no special permit was is to be expected that their presence on the footage will induce required since camera traps do not disturb the animals. bigger changes, compared to other smaller animals, e.g. small Permissions to access the camera trap locations were granted by rodents and birds, or movement of water and/or vegetation Limburgs Landschap (1 location), Natuurpunt (5 locations), nv De registered in the recordings. In the following, we propose and Scheepvaart (1 location) and Steengoed (5 locations). evaluate two different ways to quantify the amount of ‘move- Cameras were attached to a nearby tree 30–90 cm above the ments’, on which the discrimination will be based. All following ground and directed at the burrow (5 locations), a running track (3 manipulations and calculations were performed in MATLAB [44] locations), or a foraging location (4 locations). The anticipated except when stated differently. passage of a beaver was never farther than 6 m away from the To start, we performed two basic manipulations of the camera (but often closer). The camera settings were standardized recordings. The first two frames were removed from each movie over all cameras as follows: Type = Video, Video Size = 7206480, because of the time stamp on the first frame and the instability of Video Length = 15s, Interval = 1s, Time Stamp = On, Video the light caused by the starting of the camera in the first two sound = On. Cameras were activated when detecting a combina- frames. Second, we averaged out small spatial and temporal tion of movement and heat. Cameras were positioned for an changes due to detection noise, movement of vegetation, water average of 35 days at each location (range = 30 to 50 days, reflections, etc. This was done by down sampling (averaging) the SD = 5.6). The sensitivity of the sensor was set to low, medium or video along both the spatial, x and y, and temporal dimension, t high, according to local circumstances. The medium sensitivity (respectively with factor 5, 5, 10). As beavers are large and move was used in most environmental settings. When cameras were rather slowly, the detection of their movement would not suffer directed to highly dynamic streams, the sensitivity was set low. from this reduction. Moreover, this operation would also improve When beavers were expected to pass rather far from the camera computational speed. and vegetation was limited, high sensitivity was selected. The To detect non-target recordings, we first quantified the amount sensitivity must be sufficiently high since we suspect that the cold of variation throughout the video. If the pixel values did not vary water in the fur of the beavers reduces the probability of detection much throughout the video, then it was unlikely that the camera by camera traps. The illumination of recordings in poor light was triggered by an animal as large as a beaver. We used two conditions (dusk, dawn or night) was assisted by infrared LEDs, different methods to quantify this amount of pixel variation and resulting in black and white recordings. Only black and white arrived at a ‘measure of frame variation’. Both were chosen since recordings were considered since it is known that beavers are they are easy to apply and differ qualitatively from each other. PLOS ONE | www.plosone.org 2 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos Table 1. Classification matrix of the recordings. Reality Non-target recordings Target recordings Classification result Non-Target recordings True Positive False Positive Target recordings False Negative True Negative 1991 videos were recorded at 12 different locations in 9 different beaver territories, in the province of Limburg, in the east of Flanders, Belgium, between 20 July 2012 and 8 October 2012. We recorded 1043 recordings of the target species, the beaver, 553 empty recordings and 395 recordings of non-target species. Every recording was classified based on D = ‘‘the amount of pixel variation’’ as target or non-target recording. The correct classification of non-target recordings was considered to be a success (True positive, TP) since these recordings can be correctly discarded. This value must be as high as possible in order to remove the maximum amount of non- target recordings. False positives (FP) were the beavers (target recordings) which were classified as non-target recordings and wrongly discarded. This number must be as small as possible since valuable data is being discarded. False negatives (FN) were non-target recordings which were classified as being target recordings. True negatives (TN) were the target recordings which were recognized as being target recordings. doi:10.1371/journal.pone.0098881.t001 In the first method, Filter 1, we considered the frame variations the total number of non-target recordings) and the false positive rate (FP-rate; the number of beaver recordings misclassified as compared to the average frame V(x,y)~SV(x,y; t)T , of which non-target, divided by the total number of beaver recordings) that every pixel (x,y) was averaged across the length of the video (along a classification based on this threshold would produce. Variation t = time). At every time step the squared distance of the frame of the threshold over the full range of D-values, resulted in the so- V(x,y; t) to the average frame is calculated, called receiver operating characteristics curve (ROC curve, [45]), from which the costs and benefits corresponding to a particular d (t)~ V(x,y; t){V(x,y) threshold could be easily assessed. This analysis was performed in x,y RStudio [46]. We also calculated the time savings generated by using a filter We used the following equation as a measure of the ‘amount of method, as the summed duration of the recordings that would not pixel variation’: have to be inspected visually. To determine the ability of the threshold to discard the intended amount of non-target videos in 2 2 spite of sampling variation (i.e. proportion of beavers discarded in D~ max d (t) { min d (t) a a future videos), a bootstrap procedure was carried out. Briefly, a subset of 500 observations were resampled from the dataset and classified according to the original 5% FP-rate. The operation was Note that for the calculation of D we subtracted min d (t) in repeated 1000 times and the FP-ratio was calculated at each step. order to compensate for the fact that different environments show Descriptive statics for the bootstrapped FP-ratio were computed different base line variations, e.g. aquatic sceneries exhibit more for the Complete, Dry and Wet datasets. variations compared to dry area and we were interested in the movements of animals against this variable background. Results The second method, Filter 2, focused on the changes between subsequent frames During 405 camera nights, the 12 cameras recorded 1991 videos, 933 recordings in dry locations and 1058 in wet locations, 2 2 d (t)~ ½ V(x,y; t){V(x,y; t{1) , with a mean of 166 recordings per location (49–296, SD = 81.9). x,y All recordings were watched and cameras registered 1043 recordings of the target species, the beaver, 553 empty recordings and is essentially different from the previous method. The and 395 recordings of non-target species. The resulting ROC- parameter D was calculated in the same way as before (see Matlab Code S1). curves are shown in Figure 1 for the two different filters, when applying them to videos recorded in different conditions. Both Both methods quantify the changes in pixel values within the filter methods discriminated between non-target and potential recording, which we consider to be a proxy for movement. If this D value is fairly low, then the recorded movements were rather target recordings, but Filter 2 performed slightly better than Filter 1, especially when lower FP-rates were tolerated (5% FP-rate small; a larger value points to increased activity during the 15 marked by a vertical dashed line). Also, irrespective of the FP-rate, second recording. These calculated values were used to discrim- inate the assumed non-target recordings from the possible target the TP-rate in Dry circumstances was always higher compared to Wet circumstances. recordings. The FP-rate that can be tolerated depends on the study goals This discrimination was done by means of a threshold: if the calculated D-value of the video was below this threshold, then we and design, and will be most decisive in what threshold can be used. In this particular study, a loss of 5% to 20% of the beaver assumed that the video was empty (a non-target recording); we did not make any inferences on the target’s presence when the footage was tolerable. A FP-rate of 5% resulted in a reduction of threshold value was exceeded. Hence, our approach was aimed at workload allowing us to remove 26% to 53% of non-target recordings (time savings 30–53 min) in dry conditions, 13% to detecting non-target recordings. Classifying a video without a beaver (empty or with other species) as a non-target video was 33% (time savings 25–55 min) in wet conditions and 18% to 42% considered a success, see Table 1. Consequently, for a particular (time savings 56–113 min) in the complete dataset. When we threshold value, one could calculate the true positive rate (TP-rate; tolerated a FP-rate of 20%, then 72% to 76% (time savings 92– the number of correctly classified non-target recordings divided by 95 min) of non-target recordings could be discarded in dry PLOS ONE | www.plosone.org 3 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos Figure 1. Possible gain (true positive rate, TP-rate) given an accepted loss (false positive rate, FP-rate). The FP-rate represents the proportion of target recordings (beavers) classified as non-target recordings. The TP-rate is the proportion of non-target recordings correctly classified as non-target. This is the proportion of non-target recordings that will be discarded correctly given a certain FP-rate. The best performing filter maximizes the TP-rate while minimizing the FP- rate. Filter 2 performs better in all environmental circumstances. The dashed diagonal represents the outcome of a random model which cannot discriminate between target and non-target recordings. The dashed vertical line represents a 5% threshold (FP-rate). Dry,10% water area in footage (5 locations, n = 933 recordings), Wet.10% water area (7 locations, n = 1058 recordings), Complete dataset is the combined Dry and Wet dataset (12 locations, n = 1991 recordings). doi:10.1371/journal.pone.0098881.g001 conditions, but only 46% to 54% (time savings 92–104 min) in wet conditions and between 59% and 65% (time savings 191–206 min) in the complete dataset. The bootstrap analysis indicated that sampling variation had a limited effect on the FP-ratio (Table 2). Within the 95% confidence interval, a minimum of 2.9% and a maximum of 7.4% of beaver images would be discarded using the 5% threshold, based on the complete dataset. Results of the bootstrap analysis indicated that sampling variation had only a limited effect in all environmental circumstances on the percentage of record- ings discarded. Discussion We show that a filter method based on changing pixel values can partly discriminate between recordings of the study species and non-target recordings. The amount of non-target recordings that can be discarded without watching the footage depends on the chosen threshold and this threshold will vary between studies. In both filter methods, it is clear that results depend on the environmental circumstances of the camera locations. It is easier to distinguish between beaver and non-target recordings in Dry circumstances because variation is lower than in Wet circum- stances. The filtering mechanism will work best on medium to large mammals, but these are also the most suitable subjects of species inventory studies by camera traps [47]. Since these filter PLOS ONE | www.plosone.org 4 June 2014 | Volume 9 | Issue 6 | e98881 Table 2. A bootstrap analysis was performed for both filter methods (Filter 1 and Filter 2) on the complete dataset (n = 1991), on the videos recorded at dry locations (n = 933) and on the videos recorded at wet locations (n = 1058). Dry Wet Complete dataset Filter 1 Filter 2 Filter 1 Filter 2 Filter 1 Filter 2 False Positive Rate Mean 0.051 0.051 0.050 0.051 0.051 0.051 Sd 0.009 0.009 0.011 0.010 0.012 0.011 2.5 percentile 0.035 0.035 0.029 0.032 0.029 0.029 97.5 percentile 0.068 0.067 0.072 0.072 0.074 0.074 A subsample of 500 videos was randomly sampled with replacement and this was repeated 1000 times. Recordings were classified based on their D-value; the thresholds were chosen to result in a 5% false positive rate in the respective datasets (see text). The new mean, standard deviation (Sd), 2.5 and 97.5 percentile are shown. doi:10.1371/journal.pone.0098881.t002 Novel Filter Mechanism for Camera Trap Videos methods are the first attempt to partially automatically process [4,48,49] it is probably not acceptable to miss a recording since a recordings, we did not aim to create a very complicated high level single recording may have a large impact on results. For this study, pattern recognition program. The current method is rather robust a loss of 5% to 20% of the beaver footage was tolerable, since to changes in camera-angle, illumination, colour or black and beavers were recorded frequently and the data were collected to white recordings, distance from the camera to the subject, size of determine average activity patterns. A reduction of the tolerated the subject and the position of the subject to the camera since they loss of target recordings would have lowered the TP-rate, resulting can be largely accounted for by choice of the threshold. in less time savings. How quickly the TP-rate decreases with The greatest differences between filter methods were observed decreasing FP-rate depends on the shape of the ROC-curve and when the tolerated FP-rate was small. When using a FP-rate of will differ between studies. Also, these filter methods are not 5%, Filter 2 resulted in a TP–rate which was more than double the suitable for species which are very hard to record because it will value of Filter 1, in all environmental circumstances. When take a very long time before enough data are collected to increasing the tolerated FP-rate, these differences became smaller determine the threshold. (Figure 1). The filter protocol is optimized for processing videos and not Discarding up to 76% of non-target images can prove to be very photographs. The same process could be used for processing timesaving when used in long term survey studies deploying large photographs. Although each picture is a still image, the number of cameras. This reduction of workload has two direct comparison of consecutive pictures (as if they were frames of a implications for camera trap studies. Since the processing of movie) makes the analysis possible. The disadvantage of this recordings takes less time, the number of cameras can be increased method is that the time between pictures can vary a lot when allowing covering a larger study area or augmenting the amount of cameras are movement-triggered, and environmental factors, like cameras used in the same study area so that patterns can be growth of plants or changing light conditions, can make it more discerned on a finer scale. Second, the sensitivity of the camera difficult to compare different photographs. However, Hamel et al. traps can be increased, allowing the cameras to react to smaller (2013) showed that using a fixed 5-min interval resulted in lower animals. This will result in an increase of the number of daily presence raw error rates compared to movement-triggered recordings, but the time spent on processing the images will still cameras and recommend opting for time-triggered cameras when be reduced because of the filtering of non-target recordings. The aiming to capture abundant species [50]. When using fixed time recordings of small (non-target) animals may be useful for other intervals, high variation is avoided, making it easier to process studies. The same database of video recordings could then be used, images when the interval between pictures is not too long. but now with the small animal as target, without having to perform However, problems can still occur when light conditions change a new field campaign. rapidly as during sunset and sunrise. When using a series of Although we show that the proposed filter method can reduce pictures that are taken with fast intervals, as a consequence of a the amount of work substantially given a certain cost (lost of single trigger event, this problem can be avoided. recordings of interest), there are limitations to this method. Camera traps are imperfect detectors and the chance of First, a number of recordings must be viewed to determine if the detecting a species decreases with decreasing species size and species of interest is recorded. Only when a sufficient number of increasing speed and is influenced by seasons [22,51]. But a recordings of the species of interest is obtained, a comparison camera can also be ‘tricked’ into recording when there are no between these and other images can be made. The necessary animals. This generates empty recordings and these are rarely amount of recordings depends highly on the studied species (size reported although they give an important indication about the and speed), other sympatric species and the environmental efficiency and expected time effort necessary to process all conditions. Larger species that are detected regularly and will be recordings. The reasons for recording empty recordings are studied for a long time by using camera traps are the most suitable diverse, moving vegetation, technical problems, moving water, subjects to discriminate from non-target recordings. Based on the sunlight reflection or time lag between passing of an animal and determined threshold, images which most likely not contain starting of the camera. These factors will vary depending on animals can be filtered out, but a distinction between different environment, season and study species. These problems are well species of similar size and speed is not feasible. known to researchers but receive only limited attention in the The filter methods are based on changing pixel values. Animals literature [52]. To reduce the amount of empty recordings, the that do not move (stay on the same location during the length of sensitivity of the sensor can be reduced. However, this also the complete recording) will most likely be discarded. It is however increases the chance of missing the study species. Although no reasonable to assume that a motionless animal was detected while published research is available, we think that sensor sensitivity arriving or will be detected while leaving the location where the must be augmented in semi-aquatic mammals when being filmed camera trap is aimed at. Animals must be of sufficient size to stand on land since they are likely to just have left the water, resulting in out from the background noise and to be recognized as a potential a cold layer around the body of the animal making it more difficult target recording. For very small species, it may be necessary to skip to detect than a similar sized dry mammal. This high sensitivity the down sampling step in both spatial directions in order to results in more false detections, explaining the high percentage of achieve an image that is detailed enough. For very fast species, the empty recordings compared to other studies 15,7% [20]. We want down sampling in the temporal dimension can be detrimental. to encourage authors to not only mention the trapping effort and However, the performance of the method depends highly on the the amount of pictures of the studied species but also the amount variation (movement) in the background. Once the variation in the of empty recordings (or recordings of non-target species) making it background is larger than the changes induced by passing species, possible to compare how different types of cameras perform in it will not be possible to distinguish between target and non-target different circumstances allowing future researchers to benefit from recordings. this knowledge. The use of the filter protocol must be considered with Although we acknowledge the limitations of our filter method, knowledge of the species, species community and the study design. we believe an important progress has been made by showing that For example: when camera traps are used to perform capture adding an extra filter step between downloading of images and the mark recapture analysis on individually recognizable individuals PLOS ONE | www.plosone.org 5 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos locations. We thank Jemp Peeters, Nol Goossens and Roger Nijssen for statistical processing can save lots of valuable time, with losing only providing valuable help in fieldwork. We thank 1 anonymous referee and a limited amount of data. John Goodrich for the helpful comments and suggestions that considerably improved the manuscript. Supporting Information Code S1 Matlab Code. Author Contributions (M) Conceived and designed the experiments: KRRS JR MB HL. Performed the experiments: KRRS JR MB. Analyzed the data: KRRS JR MB. Wrote Acknowledgments the paper: KRRS JR MB HL. We would like to thank Limburgs Landschap, Natuurpunt, nv De Scheepvaart and Steengoed for granting permission to access camera trap References 1. O’Connell AF, Nichols JD, Karanth UK (2011) Camera Traps in Animal 25. Ariefiandy A, Purwandana D, Seno A, Ciofi C, Jessop TS (2013) Can Camera Ecology Methods and Analyses. New York: Springer. Traps Monitor Komodo Dragons a Large Ectothermic Predator? PLoS One 8: 2. De Almeida Jacomo AT, Silveira L, Diniz-Filho JAF (2004) Niche separation 1–8. between the maned wolf (Chrysocon brachyurus), the crab-eating fox (Dusicyon thous) 26. O’Brien TG, Kinnaird MF (2008) A picture is worth a thousand words: the and the hoary fox (Dusicyon vetulus) in central Brazil. J Zool 262: 99–106. application of camera trapping to the study of birds. Bird Conserv Int 18: 144– 3. Head JS, Robbins MM, Mundry R, Makaga L, Boesch C (2012) Remote video- 162. camera traps measure habitat use and competitive exclusion among sympatric 27. Seki S-I (2010) Camera-trapping at artificial bathing sites provides a snapshot of chimpanzee, gorilla and elephant in Loango National Park, Gabon. J Trop Ecol a forest bird community. J For Res 15: 307–315. 28: 571–583. 28. Rovero F, Zimmermann F, Berzi D, Meek P (2013) ‘‘ Which camera trap type 4. Hariyadi ARS, Priambudi A, Setiawan R, Daryan D, Yayus A, et al. (2011) and how many do I need?’’ A review of camera features and study designs for a Estimating the population structure of Javan rhinos (Rhinoceros sondaicus) in Ujung range of wildlife research applications. Hystrix, Ital J Mammology 24: 148-156. Kulon National Park using the mark- recapture method based on video and 29. Harris G, Thomson R, Childs JL, Sanderson JG (2010) Automatic storage and camera trap identification. Pachyderm: 90–99. analysis of camera trap data. Bull Ecol Soc Am: 352–360. 5. Karanth KU, Nichols JD, Kumar NS, Hines JE (2006) Assessing tiger 30. Daims H, Lu ¨ cker S, Wagner M (2006) Daime, a Novel Image Analysis Program population dynamics using photographic capture-recapture sampling. Ecology for Microbial Ecology and Biofilm Research. Environ Microbiol 8: 200–213. 87: 2925–2937. 31. Pennekamp F, Schtickzelle N (2013) Implementing image analysis in laboratory- 6. Royle AJ, Nichols JD, Karanth UK, Gopalaswamy AM (2009) A hierarchical based experimental systems for ecology and evolution: a hands-on guide. model for estimating density in camera-trap studies. J Appl Ecol 46: 118–127. Methods Ecol Evol 4: 483–492. 7. Karanth UK, Nichols JD (1998) Estimation of tiger densities in India using 32. Vogt A, Cholewinski A, Shen X, Nelson SG, Lazo JS, et al. (2009) Automated photographic captures and recaptures. Ecology 79: 2852–2862. image-based phenotypic analysis in zebrafish embryos. Dev Dyn 238: 656–663. 8. Rovero F, Marshall AR (2009) Camera trapping photographic rate as an index 33. Addink EA., De Jong SM, Davis SA., Dubyanskiy V, Burdelov LA., et al. (2010) of density in forest ungulates. J Appl Ecol 46: 1011–1017. The use of high-resolution remote sensing for plague surveillance in Kazakhstan. 9. Rowcliffe MJ, Field J, Turvey ST, Carbone C (2008) Estimating animal density Remote Sens Environ 114: 674–681. using camera traps without the need for individual recognition. J Appl Ecol 45: 34. Vogt P, Riitters KH, Iwanowski M, Estreguil C, Kozak J, et al. (2007) Mapping 1228–1236. landscape corridors. Ecol Indic 7: 481–488. 10. Marnewick K, Funston PJ, Karanth KU (2008) Evaluating camera trapping as a 35. Sundaresan SR, Riginos C, Abelson ES (2011) Management and Analysis of method for estimating cheetah abundance in ranching areas. South Camera Trap Data; Alternative Approaches (Response to Harris et al., 2010). African J Wildl Res 38: 59–65. Bull Ecol Soc Am: 188–195. 11. Prasad S, Pittet A, Sukumar R (2009) Who really ate the fruit? A novel approach 36. Fegraus EH, Lin K, Ahumada JA, Baru C, Chandra S, et al. (2011) Data to camera trapping for quantifying frugivory by ruminants. Ecol Res 25: 225– acquisition and management software for camera trap data: A case study from the TEAM Network. Ecol Inform 6: 345–353. 12. O’Brien TG, Baillie JEM, Krueger L, Cuke M (2010) The Wildlife Picture 37. Yu X, Wang J, Kays R, Jansen P a, Wang T, et al. (2013) Automated Index: monitoring top trophic levels. Anim Conserv 13: 335–343. identification of animal species in camera trap images. EURASIP J Image Video 13. Bridges AS, Vaughan MR, Klenzendorf S (2004) Seasonal variation in Process 52. American black bear Ursus americanus activity patterns: quantification via remote 38. Scheibe KM, Eichhorn K, Wiesmayr M, Schonert B, Krone O (2007) Long- photography. Wildlife Biol 4: 277–284. term automatic video recording as a tool for analysing the time patterns of 14. Di Bitetti MS, Paviolo A, De Angelo C (2006) Density, habitat use and activity utilisation of predefined locations by wild animals. Eur J Wildl Res 54: 53–59. patterns of ocelots (Leopardus pardalis) in the Atlantic Forest of Misiones, 39. Hariyadi AR, Setiawan R, Yayus DA, Purnama H (2010) Preliminary behaviour Argentina. J Zool 270: 153–163. observations of the Javan rhinoceros (Rhinoceros sondaicus) based on video trap 15. Tobler MW, Carrillo-Percastegui SE, Powell G (2009) Habitat use, activity surveys in Ujung Kulon National Park. Pachyderm: 93–99. patterns and use of mineral licks by five species of ungulate in south-eastern 40. Staller EL, Palmer WE, Carroll JP, Thornton RP, Sisson DC (2005) Identifying Peru. J Trop Ecol 25: 261. Predators At Northern Bobwhite Nests. J Wildl Manage 69: 124–132. 16. Janec ˇka JE, Munkhtsog B, Jackson RM, Naranbaatar G, Mallon DP, et al. 41. Sharpe F, Rosell F (2003) Time budgets and sex differences in the Eurasian (2011) Comparison of noninvasive genetic and camera-trapping techniques for beaver. Anim Behav 66: 1059–1067. surveying snow leopards. J Mammal 92: 771–783. 42. Schneider C a, Rasband WS, Eliceiri KW (2012) NIH Image to ImageJ: 25 17. Srbek-Araujo AC, Chiarello AG (2005) Is camera-trapping an efficient method years of image analysis. Nat Methods 9: 671–675. for surveying mammals in Neotropical forests? A case study in south-eastern 43. Macdonald DW (2001) Beavers. The new encyclopedia of mammals. Oxford: Brazil. J Trop Ecol 21: 121–125. Oxford University Press. pp. 590–595. 18. O’Brien TG, Kinnaird MF, Wibisono HT (2003) Crouching tigers, hidden prey: 44. Matlab (2012). Version R2012b. 8.0.0.783. The MathWorks Inc. Natick Sumatran tiger and prey populations in a tropical forest landscape. Anim Massachusetts. Conserv 6: 131–139. 45. Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett 27: 19. Silveira L, Ja ´ como AT a., Diniz-Filho JAF (2003) Camera trap, line transect 861–874. census and track surveys: a comparative evaluation. Biol Conserv 114: 351–355. ¨ 46. Rstudio (2012). RStudio: Integrated development environment for R (Version 20. Can OE, Togan I (2009) Camera trapping of large mammals in Yenice Forest 0.96.122) [Computer software]. Boston, MA. Available http://www.rstudio. Turkey local information versus camera traps. Oryx 43: 427–430. org/. Accessed 2014 Jan 10. 21. Tobler MW, Carrillo-Percastegui SE, Leite Pitman R, Mares R, Powell G (2008) 47. Kelly MJ (2008) Design, evaluate, refine: camera trap studies for elusive species. Further notes on the analysis of mammal inventory data collected with camera Anim Conserv 11: 182–184. traps. Anim Conserv 11: 187–189. 48. Soisalo MK, Cavalcanti SMC (2006) Estimating the density of a jaguar 22. Rowcliffe MJ, Carbone C, Jansen PA, Kays R, Kranstauber B (2011) population in the Brazilian Pantanal using camera-traps and capture–recapture Quantifying the sensitivity of camera traps: an adapted distance sampling sampling in combination with GPS radio-telemetry. Biol Conserv 129: 487–496. approach. Methods Ecol Evol 2: 464–476. 23. Olson ER, Marsh RA, Bovard BN, Randrianarimanana LHL, Ravaloharima- 49. Carbone C, Christie S, Conforti K, Coulson T, Franklin N, et al. (2001) The use nitra M, et al. (2012) Arboreal camera trapping for the Critically endangered of photographic rates to estimate densities of tigers and other cryptic mammals. greater bamboo lemur Prolemur simus. Oryx 46: 593–597. Anim Conserv 4: 75–79. 24. Tan CL, Yang Y, Niu K (2013) Into the night: camera traps reveal nocturnal 50. Hamel S, Killengreen ST, Henden J-A, Eide NE, Roed-Eriksen L, et al. (2013) activity in a presumptive diurnal primate, Rhinopithecus brelichi. Primates 54: 1–6. Towards good practice guidance in using camera-traps in ecology: influence of PLOS ONE | www.plosone.org 6 June 2014 | Volume 9 | Issue 6 | e98881 Novel Filter Mechanism for Camera Trap Videos sampling design on validity of ecological inferences. Methods Ecol Evol 4: 105– 52. Swann DE, Kawanishi K, Palmer J (2011) Evaluating types and features of camera traps in ecological studies: a guide for researchers. In: O’Connell AF, 51. Tobler MW, Carrillo-Percastegui SE, Leite Pitman R, Mares R, Powell G (2008) Nichols JD, Karanth KU, editors. Camera Traps in Animal Ecology:Methods An evaluation of camera traps for inventorying large- and medium-sized and Analyses. New York: Springer. pp. 27–43. terrestrial rainforest mammals. Anim Conserv 11: 169–178. PLOS ONE | www.plosone.org 7 June 2014 | Volume 9 | Issue 6 | e98881

Journal

PLoS ONEPubmed Central

Published: Jun 11, 2014

There are no references for this article.