With the development of precision livestock farming, in order to achieve the goal of fine management and improve the health and welfare of dairy cows, research on dairy cow motion monitoring has become particularly important. In this study, considering the problems surrounding a large amount of model parameters, the poor accuracy of multi-target tracking, and the nonlinear motion of dairy cows in dairy farming scenes, a lightweight detection model based on improved YOLO v11n was proposed and four tracking algorithms were compared. Firstly, the Ghost module was used to replace the standard convolutions in the YOLO v11n network and a more lightweight attention mechanism called ELA was replaced, which reduced the number of model parameters by 18.59%. Then, a loss function called SDIoU was used to solve the influence of different cow target sizes. With the above improvements, the improved model achieved an increase of 2.0 percentage points and 2.3 percentage points in mAP@75 and mAP@50-95, respectively. Secondly, the performance of four tracking algorithms, including ByteTrack, BoT-SORT, OC-SORT, and BoostTrack, was systematically compared. The results show that 97.02% MOTA and 89.81% HOTA could be achieved when combined with the OC-SORT tracking algorithm. Considering the demand of equipment in lightweight models, the improved object detection model in this paper reduces the number of model parameters while offering better performance. The OC-SORT tracking algorithm enables the tracking and localization of cows through video surveillance alone, creating the necessary conditions for the continuous monitoring of cows.

This study proposed an online early lameness detection method for dairy cow health management to overcome the inability of wearable sensor-based methods for online detection and low sensitivity to early lameness. Wearable IMU sensors collected acceleration data in stationary and moving states; a threshold discrimination module using variance of motion-direction acceleration was designed to distinguish states within 2 s, enabling rapid data screening. For moving-state windowed data, the InceptionTime network was modified with YOLOConv1D and SeparableConv1D modules plus Dropout, which significantly reduced model parameters and helped mitigate overfitting risk, enhancing generalization on the test set. Typical gait features were fused with deep features automatically learned by the network, enabling accurate discrimination among healthy, mild (early) lameness, and severe lameness. Results showed that the online detection model achieved 80.6% dairy cow health status detection accuracy with 0.8 ms single-decision latency. The recall and F1 score for lameness, including early and severe cases, reached 89.11% and 88.93%, demonstrating potential for early and progressive lameness detection. This study improves lameness detection efficiency and validates the feasibility and practical value of wearable sensor-based gait analysis for dairy cow health management, providing new approaches and technical support for monitoring and early intervention on large-scale farms.