An Outdoor 3D Object Detection Method Based on Virtual Point Cloud Generation and Category Awareness Feature Enhancement

In outdoor 3D object detection, the utilization of point cloud data still faces challenges, such as the missing point cloud data caused by object occlusion and the abundant cluttered information in complex outdoor backgrounds, which both interfere with the network's precise determination of object location and size. We present a novel outdoor 3D object detection method based on Virtual Point Cloud Generation (VPCG) and Category Awareness Feature Enhancement (CAFE) to effectively address these challenges. VPCG votes from surface voxels to object center, thereby building virtual central voxels and using sparse convolution to populate the resulting point cloud features. CAFE clusters the voxels and enhances their features by performing TRANSFORMs according to the class they belong to. To validate the effectiveness of our method, comparative experiments on the KITTI public dataset were conducted for detecting Car, Pedestrian and Cyclist. Compared to the base network called 3D Cascade RCNN, the mAP performance improvements for these three objects are +0.44, +1.38, and +0.93 at the average difficulty level, respectively.