Abstract: This paper aims to explore the application of convolutional neural networks in line patrol tasks of humanoid robots. By dividing the traditional convolution operation into deep convolution and point-by-point convolution, the number of parameters of the model is greatly reduced, and attention mechanism is introduced to improve the accuracy of the track recognition model. Firstly, the image data is preprocessed to remove the interference of noise points, and then input into the improved convolutional neural network model. Through the combination of deep convolution and point-by-point convolution, the feature information in the image can be better captured. The introduction of the attention mechanism helps to enhance the attention of the network to the key areas and further improve the accuracy of the patrol. This method has been verified in the Jiangsu University Student Robot Competition and achieved good results. Compared with the traditional methods, this method shows higher accuracy and stability in the patrol task, it can autonomously identify three cases of going straight, turning left, and turning right; the average correct recognition rate can reach 95.86%, and the operation is reliable. These research results are of great significance to further promote the development of humanoid robots in practical applications.