Prior Distribution and Perception Radius Are Important: Learning 3-D Object Detector for Autonomous Driving

This article proposes an efficient 3-D object detection algorithm for LiDAR point clouds in the field of autonomous driving. Currently, despite significant performance improvements in 3-D LiDAR object detection methods, there are still problems that have impact on 3-D detection accuracy. Typically, since the sampling process and grouping process introduced background points, the final regression step in the detection task would be greatly affected. To address this problem, a method named PR-SSD is proposed to retain more important foreground points during the training pipeline, improving detection accuracy. Specifically, the proposed approach utilizes prior information to guide semantic sampling, referred to as prior distributed semantic sampling (PDSS). This mechanism encourages the network to prioritize foreground targets in regression. In addition, a module named radius-aware attention multiscale grouping (RAMSG) is designed to dynamically reassign grouping weights for multiscale features. In other words, the network has adaptive detection capabilities for targets of different scales. Furthermore, PR-SSD is a single-stage detector which can be trained end-to-end. Finally, extensive experiments and evaluations on large-scale benchmarks KITTI demonstrate that our proposed method significantly enhanced 3-D object detector. In KITTI testing set, PR-SSD achieves 89.69%, 45.08%, and 80.01% detection performance for cars, pedestrians, and cyclists.