Autonomous navigation is required for high recognition accuracy, real-time performance, and environmental adaptability under complex ground textures and ambiguous path boundaries in orchards. In this study, a lightweight improved U-Net path recognition model, named MEG-UNet, was proposed to extract navigation lines for multi-scenario orchards. A navigation line generation algorithm was developed to incorporate with a traversability determination mechanism. In the MEG-UNet model, MobileNetV4 was first adopted as the encoder backbone to reduce computational complexity for feature extraction. Haar wavelet downsampling was introduced into the deep feature extraction to enhance the perception of global path structures and continuity. Subsequently, Ghost Convolution was employed in the decoder to further reduce the computational cost for effective feature reconstruction. Moreover, a spatial-channel collaborative efficient local attention (SCC-ELA) mechanism was embedded into the skip connection module. Spatial and channel attention features were integrated to improve the path segmentation accuracy in a multi-scenario orchard. The mechanism also guided the network to focus on critical path regions. A navigation line generation algorithm was designed with a traversability decision. A binary path image was obtained to extract the traversable region at the bottom of the image. The path width was estimated after extraction. The path traversability was determined by comparing the path width with the dimensions of agricultural machinery. Once the path was determined to be traversable, edge detection was performed to generate a midpoint sequence. Subsequently, B-spline curve fitting was applied to generate the final navigation line. Furthermore, a pipeline parallel processing architecture was introduced using a dual-buffer mechanism in collaborative processing between path recognition and navigation line generation. Experiments were conducted in multiple orchards, including plain apple orchards, hilly apple orchards, vineyards, and greenhouse strawberry orchards. The path recognition model achieved higher segmentation accuracy than classical semantic segmentation and YOLO-series models. Compared with the baseline ResNet50-UNet model, the mean Intersection over Union (mIoU) and mean Pixel Accuracy (mPA) improved by 5.27 percentage points and 2.96 percentage points, respectively, among different orchard scenarios. In the two-apple orchards, the mIoU values reached 93.06% and 92.37%, and the mPA values reached 96.85% and 95.97%, respectively. In vineyard and strawberry orchards, the mIoU values were 92.03% and 91.80%, while the mPA values were 96.32% and 96.05%, respectively. Only 8.32 million parameters were able to fully meet the requirements of the lightweight deployment. In addition, the navigation line generation algorithm achieved the maximum traversability determination accuracy of 98%, and the navigation line extraction completeness exceeded 90%. The average angular error of the B-spline-fitted navigation lines ranged from 0.37° to 2.45°, the maximum angular error was 3.26°, and the maximum distance error was within 7.52 cm. The overall algorithm operated stably in both video-stream processing experiments and field tests, with the processing frame rate above 25 frames per second. The MEG-UNet improved path segmentation accuracy in complex orchards, indicating its low computational cost. Combined with the traversability determination mechanism and B-spline-based navigation line generation, reliable path recognition and navigation guidance can be expected for autonomous agricultural equipment. The real-time performance and environmental adaptability can also provide effective technical support for visual navigation in multi-scenario orchard environments.

Different varieties of the same fruit tree can be classified as a fine granularity. Most them, only a single organ of fruit cannot fully extract the overall feature of the plant. It is also difficult to further improve the accuracy of recognition. Multi-organ features can be expected to recognize the fruit tree in current research. However, the inter-organ features can be interfered with each other in the multi-organ feature recognition. It is still challenging to accurately locate and distinguish different organ features. In this study, an organ segmentation (RSA-YOLOv8s) was proposed for the fruit and leaf using YOLOv8s target detection. The three parts of the model were improved, namely Backbone, Neck, and Head. A cross-stage local residual (Residual CSPLayer 2Conv, RC2) module was also designed in the Backbone part. The sensory field of each network layer was used to construct the hierarchical class residual links in each network layer. The dense and small target features were fully extracted after that. The Scale Spatial Pyramid Pooling (SSPP) module was designed in the Neck part. The spatial and scale information of the network were integrated to introduce 3D convolution. The higher-order and multi-scale features were extracted to detect the multi-scale targets. Furthermore, the Asymmetric decoupling detection head (ADDH) module was designed in the Head part. The regression of the model was realized to classify the features using an asymmetric structure. As such, the classification was focused on the center content, while the regression was the edge information. A total of 950 images of 17 fruit tree varieties were selected from the PlantCLEF2022 public dataset. The experimental results show that the precision rate, recall rate, F1 value and average precision of the RSA-YOLOv8s model were 83.2%, 87.9%, 85.5%, and 90.2%, respectively, in the recognition of fruit and leaf organ, which were 5.6, 6.6, 6.1 and 6.7 percentage points higher than those of the original, respectively; By contrast, the precision rate, recall rate, F1 value and mean precision were 88.2%, 91.7%, 89.9% and 94.3%, respectively, in the recognition of single organ, which were 6.4, 6.6, 6.5 and 7.2 percentage points higher than those of the original, respectively; The precision rate, recall rate, F1 value and the mean precision were 78.2%, 84.1%, 81.0% and 86.1%, respectively, in the recognition of leaf single organ, which were 4.8, 6.6, 5.6 and 6.2 percentage points higher than those of the original, respectively. In addition, the improved model was also applied into the 20 date varieties with a total of 3 881 images. A recognition accuracy of 99.1% was achieved, which was 6.6% higher than that of the original model. The finding can provide a strong reference for the organ segmentation and variety recognition using multi-organ features, particularly for common fruit trees.