A multi-attention guided convolutional network for real-time unstructured road segmentation and autonomous navigation in greenhouse environments

Semantic segmentation of unstructured greenhouse roads presents numerous challenges, including high model complexity, limited scene-specific data, and practical deployment constraints. To address the issue of model complexity, a multi-attention guided convolutional network for segmentation (MAGCN-Seg) is proposed. The architecture incorporates a dual-path heterogeneous convolution module to improve directional structural modeling and global contextual perception, and a multi-scale convolutional attention module (MSCAM) to enhance multi-scale semantic representation while suppressing background interference. To mitigate the scarcity of greenhouse-specific data, a monocular-vision-based greenhouse unstructured road dataset (GURD) was designed, and a random cropping-patching (RCP) data augmentation strategy was used to enhance model generalization. Experimental results using the GURD, off-road freespace detection (ORFD), and cavs traversability (CaT) mixed datasets indicated that MAGCN-Seg outperformed state-of-the-art methods, with improvements of 1.3%, 1.4%, and 1.9% in mean intersection over union (mIoU) and 1.1%, 1.7%, and 5.3% in mean pixel accuracy (MPA), respectively. These results confirm the effectiveness of the proposed approach in complex unstructured road scenarios. To validate its practical applicability, an end-to-end vision-control closed-loop system was developed and deployed in 0.7-m wide greenhouse corridors. The system exhibited real-time inference at 27 frames per second (FPS) on the Jetson AGX Orin platform and maintained stable navigation with a crop clearance distance of 0.02 m. The lateral deviation was controlled at <0.01 m, and the heading error remained below 0.25°, effectively compensating for the limited operational radius of inspection manipulators and demonstrating strong engineering feasibility and robustness.