Recognition of big mammal species in airborne thermal imaging based on YOLO V5 algorithm

Abstract

Unmanned aerial vehicle (UAV) technology, artificial intelligence, and the relevant hardware can be used for monitoring wild animals. However, existing methods have several limitations. Therefore, this study explored the monitoring and protection of Amur tigers and their main prey species using images from UAVs by optimizing the algorithm models with respect to accuracy, model size, recognition speed, and elimination of environmental interference. Thermal imaging data were collected from 2000 pictures with a thermal imaging lens on a DJI M300RTK UAV at the Hanma National Nature Reserve in the Greater Khingan Mountains in Inner Mongolia, Wangqing National Nature Reserve in Jilin Province, and Siberian Tiger Park in Heilongjiang Province. The YOLO V5s algorithm was applied to recognize the animals in the pictures. The accuracy rate was 94.1%, and the size of the model weight (total weight of each model layer trained with the training set) was 14.8 MB. The authors improved the structures and parameters of the YOLO V5s algorithm. As a result, the recognition accuracy rate became 96%, and the model weight was 9.3 MB. The accuracy rate increased by 1.9%, the model weight decreased by 37.2% from 14.8 MB to 9.3 MB, and the recognition time of a single picture was shortened by 34.4% from 0.032 to 0.021 s. This not only increases the recognition accuracy but also effectively lowers the hardware requirements that the algorithm relies on, which provides a lightweight fast recognition method for UAV-based edge computing and online investigation of wild animals.