Towards Automated Monitoring of Animal Movement using Camera Networks and AI

We partnered with the Arizona Game and Fish Department to use computer vision techniques to enhance road ecology studies. Various structures, including overpasses, underpasses, escape ramps, and slope jumps have been constructed in order to facilitate animal movement across major highways and to mitigate animal-vehicle collisions. The successful functioning of these structures is monitored by placing between one and nine camera traps on each structure in order to capture how wildlife interacts with it. There are over 50 camera traps deployed across Arizona and a few neighboring states, resulting in tens of thousands of images being collected every 6-8 weeks. Our goal is to increase the efficiency of image processing and eliminate human error/bias by leveraging deep neural networks (Mask RCNN). Our results so far include an 88% detection accuracy and a 40% classification accuracy for 5 labeled species. Future work will focus on improving detection and classification, increasing the number of species we can identify, and identifying sex, age, and direction of travel. Longer term goals involve building a “smart” camera network to do real-time image processing on all of these structures.