Multi-attentional Action Recognition Method for Captive Giant Panda based on Surveillance Video

The actions of giant pandas are important indicators of their physical and mental health, and understanding their actions helps to improve animal welfare and refine the management. In this paper, we present a Multi-Attentional Action Recognition (MAAR) method, which includes spatial-temporal, motion and channel attentions, for giant pandas based on surveillance videos. The Region of Interest (ROI) video clips, which are generated by YOLOv7 model, are processed by our MAAR model to identify action labels. We also created a giant panda action recognition dataset based on the surveillance video, which includes 7 types of actions with 49,689 video clips with each label. Experimental results indicate that our proposed method achieves a mean Average Precision (mAP) of 81.75%, which outperforms ACTION-Net with a 3.58% improvement.