C-DHV: A Cascaded Deep Hough Voting-Based Tracking Algorithm for LiDAR Point Clouds

Abstract

A LiDAR-based 3-D object tracking system has been widely used in various scenarios such as autonomous driving and video surveillance, as it provides real-time and accurate object locations. Existing 3-D object tracking algorithms have achieved success by employing deep Hough voting to generate 3-D proposals. However, only one-stage voting adopted to generate 3-D proposals leads to inaccurate localization and degraded performance in complex scenarios with substantial background distractors and drastic appearance change. In this article, we propose a novel cascaded deep Hough voting (C-DHV) algorithm, which employs multistage voting to iteratively refine the 3-D proposals. Specifically, in each voting stage, the geometric locations and features of 3-D proposals are refined, which provides better initialization for the next voting stage. To improve the discriminative ability of C-DHV, the hierarchical features are fully leveraged by a feature transfer module to guide each voting stage, which enables to fuse the deep-layer features into low-level voting stage. Besides, a transformer-based feature clustering module is developed to adaptively aggregate features of 3-D proposals delivered from multistage voting, which promotes the prediction of the most accurate proposal as the final tracking result. Extensive experiments on challenging KITTI, NuScenes, and Waymo Open Dataset show that our C-DHV achieves competitive performance compared to state-of-the-art methods and significantly outperforms the one-stage voting counterpart.