BEVFusion-Based Multimodal Perception Optimization: Dynamic Spatial Adaptation and Edge-Aware Enhancement for Autonomous Driving

Multimodal fusion perception can achieve more robust environment understanding in complex traffic scenes by integrating the advantages of heterogeneous sensors. Existing BEVFusion algorithms suffer from inefficient modal interaction and poor edge segmentation. To address these problems, this paper proposes a dynamic cross-modal fusion mechanism based on an adaptive spatial weighting algorithm, and also designs an edge perception module. Firstly, the spatial attention mechanism is used to dynamically allocate the feature fusion weights of the camera and LiDAR, which solves the problem of unbalanced modal contribution in complex scenes with traditional fixed weight fusion. Then, the segmentation task and the edge detection task are jointly optimized to enhance the sensitivity of the model to the boundary region by edges. Experiments show that the adaptive spatial perspective fusion module and the edge perception enhancement module improve the mIoU of segmentation by 1.3 % and the mIoU of drivable region segmentation by 1.39 % on the NuScenes dataset.