State-of-the-art 3D object detectors are designed based on datasets with sparse and single-echo point cloud information. However, with recent advancements in LiDAR sensing, it is of great significance that we understand how richer point cloud information can be leveraged to boost performance of 3D object detectors. In this thesis, we push the limit of 3D object detection by enriching point cloud in three ways: capturing multiple reflection points in each beam instead of one; capturing an additional ambient value (IR sunlight reflected) corresponding to each point; and increasing point cloud density. Specifically, based on point cloud with ambient information collected by a prototype LiDAR, we propose a multi-view segmentation detection fusion framework that enhances metric-preserving yet sparse voxel feature learning by dense observations from perspective view. The proposed framework is shown to noticeably improve pedestrian detection accuracy. Also, based on multi-echo point cloud data collected from a prototype single-photon LiDAR with increased fill factor, we significantly boost performance of state-of-the-art detectors by introducing multiple points per beam. Lastly, by leveraging a synthetic dataset, we observe notable improvements in detection accuracy when point cloud density is increased. Our results show that with proper model design, 3D object detection will benefit greatly from enriched point cloud information, which calls for new benchmarks based on more advanced LiDAR sensors.

56 pages

Thesis Committee:
Kris M. Kitani (Chair)
Matthew P. O'Toole

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]