State-of-the-art segmentation models excel in urban structured road segmentation but prove inadequate for highly unstructured off-road terrains. In this work, we propose a structure-aware lightweight bilateral segmentation network (SAL-BSNet) specifically designed for unstructured off-road environments. Our network consists of three main modules, the dual-path structure-aware lightweight convolution (DPSA-Lite Conv) module adopts dynamic snake convolution to capture elongated tubular features of roads and trees. The dual pooling dilated attention module (DPDAM) integrates attention mechanisms to enhance recognition capability for small objects and fine-grained features. The adaptive feature fusion module (AFFM) achieves adaptive multi-scale feature fusion through learnable weighting, obtaining precise boundary delineation. SAL-BSNet only contains 5.81M parameters and 15.73G FLOPs but yields 88.73% mIoU at 215 FPS on the RUGD val set. The model has been successfully deployed into a drivable path perception system, which realizes autonomous path planning and obstacle avoidance for unmanned vehicles based on the results of the road segmentation, and the demonstration verifies the effectiveness of the method proposed in the paper.