An algorithm for keypoint detection in mice based on improved YOLOv8n-pose

This research introduces an improved YOLOv8n-pose method aimed at overcoming challenges in identifying keypoints, particularly those obscured by limb overlap, as well as simplifying the model for more effective mice behavior studies. The integration of the DCNv3 convolution module, which enhances multi-scale feature capture, with the coordinate attention mechanism (CA), which enhances the sensitivity of keypoint localization, and with the weighted bidirectional feature pyramid network (BiFPN), which optimizes the efficiency of feature fusion, results in the construction of a lightweight and high-precision detection model. The experimental findings revealed that the improved model achieved an average mean accuracy (mAP) of 90.6%, signifying a 2.1% increase compared to the original YOLOv8n-pose. The model's parameter size was reduced by 9.1% to 3.0M, and its floating-point operation size was decreased by 5.7%. he proposed algorithm achieves higher detection accuracy while maintaining lightweight characteristics, effectively addressing issues such as data loss and false negatives resulting from self-occlusion. This offers robust technological support for automated behavior analysis.