Abstract:

Ecological sensing terminals represented by infrared cameras provide massive amounts of image and video data for wildlife monitoring research.To improve the problems of low timeliness and limited processing ability in manual recognition of massive data,and to solve the uncertainty of object detection models in practical scenarios affected by multiple factors such as complex backgrounds,multiple targets,light and dark,a wildlife object detection dataset was established using leopard,adult male bharal,and non-adult male bharal as examples.Four classic object detection models,Faster R-CNN,SSD,YOLOv5,and YOLOv8,were compared and analyzed in terms of detection accuracy,detection speed,and detection effectiveness in actual scenarios.The results show that the detection effect and speed of YOLOv5 and YOLOv8 are overall better than Faster R-CNN and SSD.1)YOLOv8 has higher detection accuracy and stronger robustness under multiple interference factors,making it more suitable for scenarios that pursue detection results;2)All four models can meet the real-time video detection needs of ecological perception terminals,but the YOLOv5 model is the lightest and has the fastest detection speed,making it more suitable for scenarios with limited computability that pursue detection speed.YOLOv5 and YOLOv8 have superior performance and are suitable for detecting wildlife video targets in practical scenarios.

Key words: deep learning, object detection, YOLOv8, wildlife video