To solve the current problems of low efficiency and low accuracy of automatic detection of cow lameness in farms， a cow lameness recognition method based on posture estimation and knee angle eigenvectors was designed. Given the random behavior of dairy cows， a cow posture estimation dataset was produced by combining the imaging characteristics of cows under different conditions such as near and far field of view scales and observation angles. The Faster RCNN convolutional neural network model was introduced into the key point detection of dairy cows to improve the reliability of lameness recognition. Taking ResNet101 network as feature extraction network， the cow posture estimation network was constructed， and the hyperparameter fine-tuning training method was used to train the migration of the network model. Based on the information of cow’s posture and key point coordinate in the video， the angle feature of the cow’s knee joint when walking were calculated， and the 1-D Convolution classification model was used to realize the cow's lameness recognition. The experimental results showed that the PCK@0.1 value of the cow posture estimation network based on ResNet101 network model can reach 0.925 0. Compared with the LSTM， Bi-LSTM， and GRU models， the accuracy of cow behavior classification and recognition of 1-D Convolution model was 97.22%， which was 5.55， 2.78 and 11.11 percentage points higher， respectively. The above results show that the proposed method has a better detection effect on cow lameness in natural environment， which can provide technical reference for intelligent breeding and management of dairy caws.