Depth Completion Bird's-Eye View for 3D Object Detection

3D target detection is achieved by fusing the different modal data of lidar and camera and has a wide range of applications in autonomous driving, robot visual perception, and other fields. The existing bird's-eye view fusion method requires discrete depth estimation in the camera branch, and then the features are mapped to the bird's-eye view space to fuse with the features of LiDAR and achieve 3D target detection. Therefore, the accuracy of depth estimation directly affects the accuracy of detection. In order to solve these problems, a depth completion bird's-eye view method was proposed to achieve multi-modal fusion for 3D target detection, named DCBEV. First, RGB images provide rich structural information, and affinity matrices are used to complete sparse LiDAR information to generate dense RGBD images as input for 3D target detection. Finally, features of different modalities are fused in the bird's-eye view space, and object detection is performed. Experimental results show that our method is effective and achieves competitive performance.