COS 63-4 - When did the [animal: 43%] cross the road? Documenting wildlife road crossings using time-lapse photography and computer vision

Abstract: Roadways within or between protected areas can be a major source of wildlife mortality and changes to roadway infrastructure, such as increases in lane number and travel speeds, can therefore be expected to have substantial conservation impacts. Data on wildlife mortality at the road-wildland interface provides valuable but limited information with which to judge the scale of impacts that may result from roadway changes, as species that cross existing roadway infrastructure safely will not show up in mortality surveys yet may be at risk under future conditions. The initial goal of this study was to document animal road crossings using time-lapse photography, in which raspberry pi cameras units were used to document animal crossings of particular sections of roads within the Area de Conservación Guanacaste (ACG) in northwest Costa Rica, a site in which future roadway changes are anticipated. In this proof-of-concept study pi cameras were placed on the road side and aimed down roads within the ACG. The cameras were set to take pictures every 2.5 seconds for 12 camera nights and 18 camera days (12 hours each). At night illumination was provided by a battery-powered low-glow (850nm) IR illuminator. The data contained 123 distinct road-crossing events, ranging from a single image frame to 25 minutes in length. At least twenty-two different animal species were observed, with the great curassow, Crax rubra, the most frequently observed species. All images were converted to video files and reviewed by eye, an extremely time-intensive process. Computational approaches to processing the data were explored. A simple motion detection algorithm was unable to identify frames containing animals, particularly during the day. Application of the MegaDetector object detection model, which is designed for camera trap images, was more successful. Including only vertebrate animals, the maximum sensitivity MegaDetector achieved was classifying 65% of animal images as animals, but with a specificity of 0.35. At a specificity of value of 0.9, 45% of animal images were identified. At a specificity of 0.99 only 38% of animal images were classified as such. The data collected revealed several limitations of this approach, including the amount of illumination required to stop animal motion at night to the geometric limitation that the number of pixels corresponding to a cross section of road decline rapidly with distance. This technique shows promise for wildland road monitoring, but there remain several technical hurdles to overcome.