Abstract: Livestock farming is ever increasing in production scale and automation in recent years. Accurate and efficient monitoring of individual livestock can greatly contribute to the growth and health status of animals. Behavioral and health information can often be involved in the individual livestock. Among them, the lameness behavior of a cow can indirectly represent the health status of its legs. Accurate leg tracking is very critical to detect the lameness of a cow. However, the existing intelligent detection of cow lameness is often used at the milking aisle of the farm. There are the occlusion, blurring, and deformation of the leg in this viewpoint. In this study, leg tracking and lameness detection were proposed for dairy cows in the occlusion environment. Firstly, the video of cows walking was recorded and then pre-processed in an actual farming setting. 50 video segments were chosen as the samples for the multi-target tracking dataset of cow legs. Each segment was featured as a walking cow with a duration of 8-20 s. The dataset included a total of 27 healthy and 23 lame cows. In the dataset samples, the training and testing sets were randomly divided in the ratio of 8:2. 40 videos for model training, and 10 videos were used for model testing. Secondly, the improved FairMOT target tracking model of the cow leg was designed to obtain the cow leg trajectory. Specifically, the coordinate attention (CA) module and Squeeze-Excitation attention (SE) module were introduced into the FairMOT Iterative Deep Aggregation (IDA) and Hierarchical Deep Aggregation (HDA), respectively. The important features of the cow's leg were focused on reducing the interference of redundant information under shape mutation and blurring along the spatial and channel directions. NSA Kalman filtering was introduced into the trajectory state for the updating phase, in order to obtain more accurate motion states of leg targets. Tracking performance was improved to reduce the measurement weights of low-confidence detection targets when the leg occlusion occurred; the AFLink linking mechanism was introduced into the tracking post-processing stage, in order to improve the tracking performance of the model. The curve-fitting of cow movement trajectory was also designed in the cow gait cycle under the occlusion environment. Finally, the individual thresholds were used to classify the phases of the cow's gait movement. Six gait features were extracted, namely, stride length, total gait period, support phase gait period, swing phase gait period, support phase share, and swing phase share. The Navie Bayes model was then utilized for the detection and classification of limping behavior. The experimental results show that the improved model shared better localization accuracy and tracking, compared with the original FairMOT model. Specifically, the MOTA (Multiple Object Tracking Accuracy) was improved by 3.79 percentage points, whereas, the MOTP (Multiple Object Tracking Precision) and IDs (Total Target ID Switching The total number of identity switches) decreased by 0.022 and 35 times, respectively. The improved model outperformed the existing mainstream multi-target tracking models. In addition, CowMOT shared the more accurate target localization more suitable for tracking the movement of a cow's leg; The classification accuracy of lameness detection was 86.67%, indicating better defection on the lameness in dairy cattle. The finding can also provide a strong reference to intelligent monitoring of cow lameness behavior.