Multi-behavior detection of group-housed pigs based on YOLOX and SCTS-SlowFast

Accurately and rapidly recognizing the behaviors of group-housed pigs plays a very important role in pig farm production management. Recognizing multiple behaviors in one scene remains a challenge. This study proposed a multi-behavior detection method of group-housed pigs based on YOLOX and SCTS-SlowFast (SC-Conv-TSSlowFast) to recognize four behaviors (Eating or Interaction with feeding though, Drinking or Interaction with drinker, Standing and Walking) and locate corresponding pig locations. Firstly, YOLOX object detection module is used to locate the locations of group-housed pigs. Secondly, SCTS-SlowFast behavior recognition module is proposed to classify the behaviors category of pigs in the located regions, in which Self-Calibrated Convolution (SC-Conv) and Temporal-Spatial (TS) attention mechanism are specially introduced to improve behavior feature extraction capability of the model. Finally, the results of two modules are combined to realize the task of multibehavior detection of group-housed pigs. To evaluate the proposed method, a multi-behavior video dataset of group-housed pigs with 420 video segments is established. This study achieved a mAP value of 80.05% for four behaviors of group-housed pigs, and counted the duration of these behaviors throughout one day from 8:00 to 16:00 as well as their corresponding changing trends. It verifies the potential and feasibility of proposed method in automatically and simultaneously monitoring and analyzing multiple typical behaviors of group-housed pigs. We shared our behavior detection dataset at https://github.com/IPCLab-NEAU/Group-housed-pigs-Multi-Beha vior-Detection for precision livestock farming research community.