Recent years, with a boundless growth in the field of autonomous driving, three-dimensional object detection technologies based on LiDAR and cameras have attracted extensively attentions. However, the LiDAR cannot obtain the complete point cloud information of the object, especially for the object that is heavily occluded. The existing multi-view detection schemes rarely process point cloud data. Thus, in this paper, on the basis of the two-stage detection network, a sampling method based on cylindrical space is proposed, combining the original point cloud and the output of the first region proposal network (RPN) to enhance the point cloud of the objects in the three-dimensional space. The second RPN will complete the accurate regression and classification of the three-dimensional bounding box by using enhanced point clouds. We solve the problem of object point cloud incomplete well. The proposed scheme is verified on the KITTI dataset. The verification results show that the method not only has good results in car three-dimensional and Bird’s Eye View (BEV) object detection, but also achieves the state of the art results in pedestrian detection, especially on heavily occluded objects.
Multi-view 3D Object Detection Based on Point Cloud Enhancement
Lect. Notes Electrical Eng.
International Conference on Intelligent Transportation Engineering ; 2021 ; Beijing, China October 29, 2021 - October 31, 2021
2021 6th International Conference on Intelligent Transportation Engineering (ICITE 2021) ; Chapter : 42 ; 476-485
2022-06-01
10 pages
Article/Chapter (Book)
Electronic Resource
English
Covariance based point cloud descriptors for object detection and recognition
British Library Online Contents | 2016
|CNN based Multi-View Object Detection and Association
IEEE | 2019
|