Map semantic understanding in obstacle-crossing scenarios is crucial for enhancing the autonomous navigation capabilities of ground mobile robots in unknown complex terrain. However, existing semantic understanding methods often suffer from issues such as noise interference, high computational costs, and low processing efficiency, which limit their ability to adapt flexibly and efficiently to real-time semantic understanding needs in unknown scenes. To address the terrain semantic understanding task in obstacle-crossing scenarios, an improved segmentation algorithm called MB-DeepLabV3+ is proposed for RiDAR point cloud semantic understanding. First, RiDAR point cloud data is used to construct an elevation raster map, with elevation values mapped to grayscale values of a 2D image, resulting in a small-sized, low-storage 2D representation. Then, the lightweight MobileNetV3 is employed as the backbone network in the DeepLabV3+ model design to reduce model complexity. Additionally, an edge-aware module (BAM) is introduced into the decoder to enhance edge features. Through supervised training, the model autonomously learns to distinguish flat regions, obstacle avoidance regions, and obstacle crossing regions. This method conserves computing resources that would otherwise be required for directly processing point cloud data. Meanwhile, the experimental results show that the improved MB-DeepLabV3+ is superior to segmentation models such as PSPNet and UNet, meets the actual needs of terrain semantic understanding tasks in obstination-crossing scenarios, and has high application value.