To solve the problem of autonomous travel and U-turn between rows for orchard visual navigation robots, a navigation line extraction method based on Mask R-CNN and a tree line extraction method based on random sample consensus (RANSAC) algorithm were proposed. Firstly, road and tree trunks were identified based on the Mask R-CNN model, and road segmentation mask and trunk bounding box coordinates were extracted. Secondly, after generating inter-row navigation lines, the improved RANSAC algorithm was used to extract the front row line of trees. Then, the distance from the coordinate point of the trunk bounding box to the front row line was calculated, and the coordinate points of the back row trunk was filtered to generate the back row line by least squares fitting. Finally, the U-turn direction can be determined by analyzing the front and back row tree lines information combined with the end of row distance and the proposed U-turn path planning method. The experimental results showed that the average segmentation accuracy and bounding box detection accuracy of the model were both 97.0% in the six orchards under different lighting, weed and weather environments. The average deviation of navigation target point extraction was within 5.3%, and the accuracy rate of tree line detection was higher than 87%. The average deviation of the vehicle position from the center of the road after the U-turn was 7.8cm. It can be proved that the proposed method can navigate effectively for visual autonomous navigation in the orchard environment.