<p>Accurate terrain perception and occupancy grid map (OGM) construction are crucial for autonomous robots in unstructured environments. This paper proposes DG-OGMNet, a real-time stereo vision-based network with depth guidance. It employs a depth-cooperative projection mechanism, fusing precise depth with camera geometry to establish a physically consistent perspective-to-BEV transformation, effectively mitigating depth ambiguity. A novel quadruple-class decoding strategy is introduced, segmenting the local OGM into flat ground, slow-moving ground (e.g., rugged terrain), obstacles, and unknown background areas. This fine-grained decomposition addresses a key gap in unstructured terrain modeling. On the KITTI dataset, DG-OGMNet significantly outperforms the single-modal method SegNeXt_L (59.25%) with an average cross-union ratio (mIoU) of 64.62%. This method achieves an IoU of 33.46% in the detection of low obstacles, which is a significant improvement compared with the comparison methods. It also achieves the best performance of 91.43% IoU in ground reconstruction tasks. Meanwhile, the network achieves a real-time inference speed of 27.3 FPS on the NVIDIA RTX 4060 GPU, with a computational complexity of only 38.55 GFLOPs and a parameter count of 14.16 MB, demonstrating excellent computational efficiency. The ablation experiment further verifies the effectiveness of each module. In cross-dataset testing, this method maintains a performance retention rate of 87.8% on the DSEC dataset, demonstrating excellent generalization ability. The ROS2 verification experiment confirms the practical value of this method in dynamic scenes, providing a new paradigm of high-precision and high-efficiency terrain perception for low-cost mobile robots.</p>