基于改进SlowFast的家猫行为识别模型OA

Domestic cats′ behavioral characteristics can directly reflect their physical and emotional wellness.Accurately recognizing their daily behaviors allows for early recognition of pet abnormalities and improves their well-being.Existing animal behavior recognition methods are mostly geared toward livestock and cannot accurately recognize domestic cat behaviors,for these behaviors are usually characterized by small motion amplitudes,quick temporal dynamics,and complex backgrounds.To address this issue,an SMT-SlowFast model was proposed,in which the temporal shift module(TSM)and the SE-SimAM(SEAM)attention module were collaboratively introduced into both the Slow and Fast pathways based on the SlowFast framework.Features of specific channels were shifted forward and backward by TSM along the temporal dimension,allowing for efficient interaction between adjacent frames and improving the modeling capability for dynamic behavioral changes over time;three-dimensional weight of channel-level attention and neuron-level importance were integrated by SEAM,resulting in highlighted key behavioral regions and effectively suppressed complex background interference.In order to validate the SMT-SlowFast model,the multi-scene domestic cat behavior dataset(MDCBD)was created.Experimental results show that,for the eight behaviors,the SMT-SlowFast model outperforms several baseline models,with an accuracy of 94.60%.Notably,SMT-SlowFast provides particularly substantial improvements in recognizing behaviors like grooming,running,and walking.This study provides an efficient and robust technical approach for non-contact monitoring of domestic cat behavior.