The skeleton extraction of cows is based on the prediction of key points of cows, which can provide important reference for detection of claudication, analysis of estrus behavior, and estimation of motion of cows through point and line reconstruction of skeleton structure of cows. Based on the partial affinity field, taking the video taken by the monitoring camera of the farm as the original data, totally 1600 images were used to train the cow skeleton extraction model, and the prediction of the key point information and partial affinity field information of the cow in the standing and walking states was realized, and the accurate extraction of the cow skeleton structure through the optimal matching connection was realized. In order to verify the performance of the model, totally 100 images of single cow and 100 images of double cows were tested. The experimental results showed that the model had a 78.90% confidence in the attitude of single target walking cows, and a 10.96 percentage points decrease in the confidence of double target walking cows compared with single target walking cows. In order to test the overall accuracy of the model, the accuracy of the model under different key points similarity OKS was calculated, and the accuracy rate was 93.40% when strict standard OKS was 0.75. Furthermore, the experimental results showed that the method can extract the cow skeleton in the video, which had high confidence and low missing rate when there was no occlusion, and the confidence was decreased when there was serious occlusion. For single target and multitarget detection, the frame processing speed of the model was 3.30f/s and 3.20f/s, respectively, and the speed was basically the same, which can lay the foundation for multitarget cow skeleton extraction. The results showed that the model can extract the skeleton information of dairy cattle accurately and could be used for the research of lameness and calculation of motion behaviors.