A top-down deep neural network for multi-dairy cows pose estimation and lameness detection

Abstract

Cow pose estimation and real-time health monitoring are important for refined herd management, improved animal welfare, and reduced passive culling rates. However, existing multi-object pose estimation methods often struggle to adapt to multi-scale objects in complex environments and typically exhibit low accuracy in detecting occluded keypoints. To address these challenges, this study proposes a top-down deep neural network for multi-dairy cows pose estimation and lameness detection, which integrates lightweight object detection, multi-scale feature fusion, and comprehensive motion feature analysis to improve the robustness under complex farm conditions. First, the real-time object detector YOLOv8n is improved by introducing the Partial Convolution (PConv) and Slim-neck modules, which improve both the efficiency and accuracy of object bounding box predictions, providing a solid foundation for the subsequent pose estimation. Second, a Path Aggregation Feature Pyramid Network (PAFPN)-based multi-scale feature fusion module is introduced as the neck network within the Real-time Multi-person Pose Estimation (RTMPose). This is further supported by a transfer learning strategy to improve keypoint localization, particularly under-occlusion and scale variation conditions. The experimental results show that the improved model achieves a mean average precision (mAP) of 95.8 %, significantly outperforming the baseline model and other existing algorithms. Seven motion features, including gait symmetry, head swing amplitude, and back curvature, were extracted in real time through pose tracking and motion trajectory analysis. These features were normalized and input into a Random Forest classifier for lameness detection. The model was evaluated on a dataset of 418 dairy cows and achieved average accuracy, sensitivity, and specificity values of 93.8 %, 94.4 %, and 97.5 %, respectively. These results demonstrate that combining multiple motion features provides a more accurate assessment of lameness.