Yak Behavior Recognition and Multi-Object Tracking Based on Improved YOLO11

In the field of animal husbandry on the Qing-hai-Tibet Plateau, current yak grazing management primarily relies on manual labor, lacking accurate recognition of yak behavioral states and health monitoring. This paper proposes a yak behavior recognition and multi-object tracking method based on an improved YOLO11n framework. First, we constructed a yak behavior recognition dataset specific to the Tibetan Plateau. Second, we enhanced the baseline YOLO11n model by replacing partial C3K2 modules in the Backbone and Neck with C3K2_DSConv modules incorporating Dynamic Snake Convolution (DSConv), enabling adaptive deformation perception for slender tubular structures and curved biological morphologies. Additionally, we integrated the C2PSA module with the EMA attention mechanism in the Backbone to create C2PSA_EMA, thereby improving multi-scale feature retention and semantic correlation of diverse features in yak behavior recognition. Finally, the optimized YOLO11n model was combined with the ByteTrack algorithm to achieve multi-object tracking for yak behavior analysis. Experimental results demonstrate that the improved YOLO11n model outperforms the baseline, with precision, recall, and mean Average Precision (mAP) increasing by $1 \%, 9.7 \%$, and 3.2 %, respectively. The refined model achieves a mAP of 83.7 %, with computational complexity and model size maintained at 7.3 GFlops and 7.5 MB, striking an optimal balance between accuracy and real-time processing.