Computer-Vision Framework for Automated Analysis of Animal Movement Ecology

Abstract

This research paper introduces a new computer-vision system that can be deployed to automatize animal movement ecology. With the help of the development of high-resolution video capture (camera-traps and drone shots) and cutting-edge object-detection and motion-tracking algorithms, the framework transforms raw images into rich trajectories and conventional movement measures (e.g., stride length, turning angle, displacement). The method was tested on a semi-natural reserve, with the video of medium-sized terrestrial mammals passed through a modular pipeline that included object detection (fine-tuned Mask -CNN), multi-object tracking (adapted sort algorithm) and geo-referenced trajectory extraction. The use of performance evaluation showed that detecting performance was 0.91 with 0.88 recall (identity-switch rate 0.07/1,000 frames), a trajectory root-mean-square error of 1.42m over manual annotated paths. It took less time by approximately 85 percent compared to manual procedures on annotation.