【摘要】 为解决智能农业机械在非结构化道路自主导航中存在的道路边界模糊、道路类型多样及夜间识别难度大等问题，基于改进DeepLabV3+提出了一种非结构化道路可行驶区域高精度识别模型。该模型以MobileNetV2替换原主干网络以降低参数量与计算复杂度，在空洞空间金字塔池化(ASPP)模块后引入坐标注意力(Coordinate Attention, CA)机制增强道路边缘与细节特征提取能力，结合迁移学习策略(Transfer Learning)与自建夜间田间道路数据集提升模型泛化性能。实验结果表明，改进模型在分割精度与推理效率上均有明显提升，整体性能优于UNet、PSPNet等主流语义分割模型，可有效实现非结构化道路可行驶区域的精准识别，能较好适配智能农业机械自主导航的需求。更多还原

【Abstract】 To address the challenges of ambiguous road boundaries, diverse road types, and difficult nighttime recognition encountered in the autonomous driving of intelligent agricultural machinery on unstructured roads, this paper proposes a high-precision drivable area recognition model based on an improved DeepLabV3+ architecture. The model employs MobileNetV2 to replace the original backbone network, thereby reducing parameter count and computational complexity. Additionally, a Coordinate Attention(CA) mechanism is introduced following the Atrous Spatial Pyramid Pooling(ASPP) module to strengthen the extraction of road edges and fine details. By integrating a transfer learning strategy with a self-built nighttime field road dataset, the model further achieves enhanced generalization performance. Experimental results demonstrate that the proposed model yields significant improvements in both segmentation accuracy and inference efficiency, outperforming mainstream semantic segmentation networks such as UNet and PSPNet. The model effectively ensures the precise recognition of drivable regions on unstructured roads, demonstrating strong adaptability to the practical requirements of autonomous navigation in intelligent agricultural machinery.更多还原