In response to the issues of accuracy and speed being difficult to balance simultaneously, as well as the weak generalization ability in the current navigation path recognition methods of fruit ridges, an optimization approach was proposed based on the U-Net model. The optimization involved integrating MobileNet-v3 Large as the backbone feature extraction network for U-Net and introducing coordinate attention at the skip connections to construct a lightweight path recognition model. Based on the drivable area segmented by this model in the inter-ridge, the edge points of the area were reshaped by using the least squares method, and further the inter-ridge navigation lines were extracted. Firstly, the model was trained on the augmented strawberry interrow dataset, and further migrated to the grape and blueberry datasets for weight fine-tuning to improve the model’s adaptability. Finally, the navigation path was identified on the corresponding verification set, and the recognition results of different models were compared visually to verify the accuracy of the model. Experimental results demonstrated that the model achieved an average intersection over union of 98.06%, 97.36%, and 98.50% for strawberry, blueberry, and grape interridge navigation path segmentation accuracy respectively, and the average pixel accuracy reached 99.13%, 98.75%, and 99.29%. The theoretical reasoning speed of the model for segmenting of RGB images was up to 19.23f/s, and the average time from image input to completed path extraction was 0.211s, meeting the requirements of real-time navigation and accuracy. A method of path extraction based on semantic segmentation was proposed, which provided a general method for the navigation of agricultural machinery equipment in interridge operation.