DGPL: A Cross-View Semantic Segmentation Perception Model in Bird's-Eye View

To address the challenges of low feature fusion efficiency, weak long-range object recognition, and high computational overhead in cross-view transformation for autonomous driving perception, we propose DGPL—an end-to-end Bird's Eye View (BEV) semantic segmentation perception model. The model simultaneously extracts multi-camera image semantic features and pixel-level depth information through a shared convolutional neural network. Innovatively, it incorporates a non-uniform learnable depth estimation module to achieve high-precision depth prediction for near-field targets while ensuring efficiency for far-field targets. A reverse grid pooling optimization module is designed to significantly accelerate the projection of 3D features into BEV space. Concurrently, a distance-aware loss function is optimized to enhance recognition capability for distant small objects. Experiments on the nuScenes dataset demonstrate that DGPL achieves state-of-the-art performance in 3D object detection with 42.5 mAP and 45.8 NDS, while outperforming mainstream methods in BEV semantic segmentation tasks. Moreover, the model requires only 9.8 million parameters and achieves an inference speed of 38.5 fps, substantially enhancing feasibility for engineering deployment.