Automated Behavior Recognition and Tracking of Group-Housed Pigs with an Improved DeepSORT Method

Pig behavior recognition and tracking in group-housed livestock are effective aids for health welfare monitoring commercial settings. However, due to demanding farm conditions, the targets pig videos heavily occluded overlapped, there illumination changes, which cause error switches of identify (ID) process decrease quality. To solve these problems, this study proposed an improved DeepSORT algorithm object tracking, contained three processes. Firstly, two detectors, YOLOX-S YOLO v5s, were developed detect classify four types behaviors including lying, eating, standing, other. Then, was reducing changes ID by improving trajectory processing data association. Finally, we established public dataset annotation pigs, with 3600 images a total from 12 videos, suitable applications. The advantage our method includes aspects. One is that association aiming at pig-specific scenarios, indoor scenes, number target objects stable. This improvement reduces enhances stability tracking. other classification information detectors introduced into In experiments detection recognition, v5s achieved high precision rate 99.4% 98.43%, recall 99% 99.23, mean average (mAP) 99.50% 99.23%, respectively, AP.5:.95 89.3% 87%. based on obtained multi-object accuracy (MOTA), (IDs), IDF1 98.6%,15, 95.7%, respectively. Compared DeepSORT, it 1.8% 6.8% MOTA IDF1, IDs had significant decrease, decline 80%. These demonstrate can achieve stable values under conditions provide scalable technical support contactless automated monitoring.