基于BEV特征融合的3D目标检测方法

Autonomous vehicles have developed rapidly in recent years. Driving safety is the most impor⁃ tant factor for autonomous vehicles， and the safety requires good perception algorithms to ensure it. The existing technical route usually chooses the BEV perspective to fuse features from different sensors， but the fusion network used in the existing research is relatively simple. Therefore， in this paper s a feature fusion network is designed to fuse BEV features from multi-sensor and multi-modality to alleviate the problem of spatial misalignment between BEV features， enhance BEV features， and improve the accuracy of 3D object detection. Considering the insufficient progress of depth prediction of image data， in this paper an image depth supervision network is also designed， which uses point clouds to generate Gaussian depth maps to directly supervise the training process of the depth pre⁃ diction network. The experimental results show that the mAP and NDS of the network on nuScenes reaches 0.669 and 0.698 respectively. The image depth predicted by this method is more continuous and the depth jump is re⁃ duced， with the edge information of BEV features clearer and the features with potential target more significant..