The cow single target tracking technology is a new technology proposed for intelligent management of dairy farms and it is the basis for the research of cow multi-objective tracking. The presence of padding in the deep network will destroy the translation invariance of the tracking model, the number of redundant parameters, and other addressing issues such as low quality of positive sample selection for tracking models, poor generalization ability of tracking models will also affect the cow tracking performance. Thus a high-performance cow single-target tracking method was proposed. Firstly, Siamese-remo model was used to extract features by improving Mobileone network to reduce the damage of tracking translation invariance by deep network to some extent, and different feature fusion parameters were preseted to train network classification and regression respectively; secondly, traditional method and go-turn method were combined to design a positive and negative sample selection strategy to improve the quality of positive sample collection; then special data enhancement was used to increase the generalization ability of the model; finally, Center-rank loss function was added to optimize the network classification and regression strategy according to the sample point location affecting confidence and IOU ranking. The experiment proved that the expected average overlap (EAO) of the cow single target tracking model in natural scenes reached 0.475, which was improved by 0.078 relative to the baseline model, and achieved better results compared with existing trackers. The number of parameters was only onetwentieth of the existing mainstream algorithms, which provided strong technical support for the subsequent cow identification and target tracking system.