Timely monitoring of cow estrus is very important in dairy cow breeding. At present, artificial estrus monitoring of dairy cows is time-consuming and laborious. Pedometer contact monitoring can easily cause stress discomfort to cows. Aiming at the problems existing in cow estrus monitoring, according to cows span behavior characteristics during oestrus, a method of cow’s oestrus behavior recognition based on convolutional neural network (CNN) was proposed. The convolution neural network was constructed to improve the network training speed by batch normalization. Max-pooling was used as the down sampling, rectified linear units (ReLU) was used as the activation function, and softmax regression classifier was used as the output layer. Through the theoretical analysis and experimental verification, the network structure and parameters of 32×32-20c-2s-50c-2s-200c-2 were designed. Through video surveillance of dairy cow activity area, 150 video segments with oestrus span behavior were extracted from 50 cows behavior videos within 6 months. The network training data of 23000 frames and the test data of 7000 frames were randomly selected from selected video segments, which were used to train and test the CNN. The results showed that the recognition accuracy of estrus behavior in dairy cows was 98.25%, the missed detection rate was 5.80%, the false recognition rate was 1.75%, and the average recognition time of single frame image detection was 0.257s. It proved that the method could realize the contactless realtime monitoring of the cow’s estrus span behavior and had a high recognition rate for cow estrus. It can significantly improve the management efficiency of large-scale farming, and had a good application prospect.