Aiming at the problems of complex image background and numerous interference factors faced by visual navigation systems in orchard environments, this paper proposes an orchard navigation line recognition method based on U-Net. Firstly, the drivable areas in the collected images are labeled using Labelme (a graphical tool for image annotation) to create an orchard dataset. Then, the Spatial Attention (SA) mechanism is inserted into the downsampling stage of the traditional U-Net semantic segmentation method, and the Coordinate Attention (CA) mechanism is added to the skip connection stage to obtain complete context information and optimize the feature restoration process of the drivable area in the field, thereby improving the overall segmentation accuracy of the model. Subsequently, the improved U-Net network is trained using the enhanced dataset to obtain the drivable area segmentation model. Based on the detected drivable area segmentation mask, the navigation line information is extracted, and the geometric center points are calculated row by row. After performing sliding window processing and bidirectional interpolation filling on the center points, the navigation line is generated through spline interpolation. Finally, the proposed method is compared and verified with U-Net, SegViT, SE-Net, and DeepLabv3+ networks. The results show that the improved drivable area segmentation model has a Recall of 90.23%, a Precision of 91.71%, a mean pixel accuracy (mPA) of 87.75%, and a mean intersection over union (mIoU) of 84.84%. Moreover, when comparing the recognized navigation line with the actual center line, the average distance error of the extracted navigation line is 56 mm, which can provide an effective reference for visual autonomous navigation in orchard environments.

To address the issues of false detections and missed detections in object detection for intelligent driving scenarios, this study focuses on optimizing the YOLOv10 algorithm to reduce model complexity while enhancing detection accuracy. The method involves three key improvements. First, it involves the design of multi-scale flexible convolution (MSFC), which can capture multi-scale information simultaneously, thereby reducing network stacking and computational load. Second, it reconstructs the neck network structure by incorporating Shallow Auxiliary Fusion (SAF) and Advanced Auxiliary Fusion (AAF), enabling better capture of multi-scale features of objects. Third, it improves the detection head through the combination of multi-scale convolution and channel adaptive attention mechanism, enhancing the diversity and accuracy of feature extraction. Results show that the improved YOLOv10 model has a size of 13.4 MB, meaning a reduction of 11.8%, and that the detection accuracy mAP@0.5 reaches 93.0%, outperforming mainstream models in comprehensive performance. This work provides a detection framework for intelligent driving scenarios, balancing accuracy and model size.