| In view of the fact that the current orchard environment visual navigation line is susceptible to interference from light and weeds, and the existing navigation line generation algorithm is too complicated and has a narrow application range, this paper proposes a method to extract orchard road navigation line based on deep learning. YOLOV3 convolutional neural network is used to extract the feature points on the image and generate the navigation line by least square fitting. 800 pictures of orchard roads collected under different conditions are used as training sets, and then tested on an independent test set consisting of 240 pictures. The overall recognition rate is 95.37%. In the environment with fewer weeds, more weeds, high light and normal light, the mean deviation of navigation line is 2.15 pixels, 2.28 pixels, 2.32 pixels and 2.41 pixels, and the mean deviation distance is 3.4 cm, 3.5 cm, 2.7 cm and 3.6 cm, respectively.