Efficacy of automated detection of motion in wildlife monitoring videos

Abstract

Documenting biodiversity and wildlife behavior is time‐consuming and potentially invasive. Remotely placed cameras often are used to increase sampling and minimize disturbance of focal animals. Such wildlife monitoring programs can entail numerous and lengthy videos requiring massive amounts of time to analyze their content. We evaluated the efficacy of a computer program, MotionMeerkat (http://benweinstein.weebly.com/motionmeerkat.html), for automated detection of motion in 21 continuous, 3‐hour‐long videos monitoring woodpecker (Picidae) nest cavities during 2016 in Washington, USA. The program produces still‐image frame captures from the video where motion was detected, files logging the video time‐stamps of each frame, and x,y coordinates within each frame where motion was detected. We compared program results with video reviews conducted by students using a standard protocol for identifying movement events and species identification. The program tagged 60% more motion events of all types, and 38% more motion events of target species (woodpeckers and potential predators), than were found by student reviews. We also conducted sensitivity analyses of the program by varying program input settings; results suggest that optimal settings depend on the video quality, environmental conditions, desired target species, and type of motion. We conclude that automated motion‐detection programs such as MotionMeerkat can play important roles in use of continuous videos for wildlife monitoring, but results need to be reviewed by the researcher. © 2019 The Wildlife Society.

Towards the fully automated monitoring of ecological communities

“…behavioural and morphological traits, abundances and distributions across multiple species) offer the opportunity to develop new predictive frameworks, which for the first time can synthetise data across ecological scales (from individuals to populations) and help developing novel early warning signals that precede population and community collapses (Cerini et al, 2022 ). Indeed, ecological forecasting is an area where automated frameworks offer significant opportunity, as the resolution of data required to develop robust predictive tools is most often impossible to obtain with non‐automated methods (Cordier et al, 2018 ; Darras et al, 2019 ; Lamprey et al, 2020 ; Marcot et al, 2019 ; Wearn & Glover‐Kapfer, 2019 ; Welbourne et al, 2015 ). Moreover, automated methods allow the acquisition of these data in real‐time, pushing ecological research from the post hoc era to one where forecasts about ecosystems fate are continually updated based on the current observed state, similar to weather forecasting (Deyle et al, 2016 ; Huang et al, 2019 ; Slingsby et al, 2020 ).…”

“…However, ecologists are also typically interested in complex biotic metrics such as the behaviours, locations and traits of individuals, as well as species abundances, distributions and interactions, which ultimately define ecological communities. Technologies such as acoustic sensors and camera traps can rapidly, remotely, non‐invasively and automatically collect high‐resolution sounds and images, thus replacing, augmenting and surpassing human sampling abilities (Cordier et al, 2018 ; Darras et al, 2019 ; Marcot et al, 2019 ; Wearn & Glover‐Kapfer, 2019 ; Welbourne et al, 2015 ). Nevertheless, processing such data into meaningful ecological measurements remains a challenging task to automate and a critical operational bottleneck (Keitt & Abelson, 2021 ).…”

Seeing wildlife behavior in a new way: Novel utilization of computer vision for focal reptile videography behavior study

“…Though not evaluated by our study due to subjectivity, the ability of MotionMeerkat to be run as a background task (not requiring active attention and time by a human like the manual review method) should not be overlooked within the context of efficiency. Worth noting is that our study question was arguably simple presence of 4 infrequently observed behavior events, requiring less time for manual review than previous studies utilizing MotionMeerkat (i.e., approximately 10 minutes for MM to review a 2-hour video recording compared to an average of 2.5 hours for a human observer per recording in Marcot et al 2019).…”

Machine learning pipeline extracts biologically significant data automatically from avian monitoring videos

“…Using an automatic data collection pipeline based on Deep Meerkat (Weinstein, 2018b), we extracted visitation rates from parental provisioning videos of house sparrows and found that automatic feed rates positively correlated with manual annotation, and can reproduce biological results, equivalent to ~800 hours (~100 8-hour workdays) of human labour work. Even though the computational time for Deep Meerkat is 1:1 (1 hour video takes ~1 hour to process; see Marcot et al, 2019), computing time is much cheaper than human labour time, especially when techniques like parallel computing were used to further speed up processing. As such, we processed a huge backlog of unprocessed videos from the Lundy sparrow system that would have been infeasible without the use of machine learning methods.…”

“…Particularly, Weinstein (2018b) developed an open-source tool named Deep Meerkat which uses convolutional neural networks (CNNs) to capture movement events from wildlife monitoring videos. Despite the name, the software was initially designed for use with a hummingbird population (Marcot et al, 2019;Weinstein, 2018b), but the software has been adapted for use in marine (Sheehan et al, 2020) and insect (Mertens et al, 2021;Pegoraro et al, 2020) systems. To the best of our knowledge, no literature exists that documents the use of the software in avian systems other than the original hummingbird population.…”