3D point cloud target detection based on pseudo segmentation for autonomous driving

Abstract

Object detection plays an important role in autonomous driving. In the past decades, many object detection methods relied on 2D images, losing spatial information due to projecting 3D space into 2D space. Recently, LiDAR has become a popular sensor for 3D point cloud target detection. This paper proposes a new RCNN detection framework based on pseudo segmentation (PS-RCNN). This model is designed to achieve accurate and efficient detection on point cloud by transmitting feature information reversely. The information transmission is supervised by the semantic segmentation task. In order to reduce the difficulty in labeling, a novel algorithm is designed to generate segmentation pseudo-labels. Experimental results conducted on KITTI Dataset and Waymo Open Dataset demonstrate that our model outperforms its counterparts for detecting small objects with a balance between accuracy and efficiency.