Fast LiDAR R-CNN: Residual Relation-Aware Region Proposal Networks for Multiclass 3-D Object Detection

Three-dimensional (3-D) object detection from Light Detection and Ranging (LiDAR) point clouds is the most challenging problem in practical 3-D scene understanding. This paper presents a fast two-stage 3-D object detection framework that jointly integrates voxel and point feature representations. Specifically, the first stage takes the voxel features from raw point clouds as inputs and then outputs bird eye’s view (BEV) feature maps and structured voxel center points. The BEV feature map and objects’ empirical sizes are used for generating 3-D proposals. The second stage extracts region pointwise features for the final object prediction using the 3-D proposals generated in the first stage. The proposed framework runs at 30 frames per second (FPS) with high performance. To improve the performance of the pedestrian class, we propose a dual-path feature module (DFM) to learn and pass features from BEV feature maps. Moreover, we propose a lightweight relation-aware module (LRAM) for sparse point clouds to enhance the attention ability of region proposal networks by exploring the relationships between pixels and between channels. On the KITTI benchmark suite, performed experiments show that the proposed LiDAR-based method achieves a new state-of-the-art on the three classes in 3-D performance (Easy, Moderate, Hard): car (92.53%, 84.70%, 82.32%), pedestrian (68.30%, 61.20%, 55.17%), cyclist (91.73%, 72.61%, 68.24%).