Cheng, Xiaolong (Southeast University), Ma, Tianxiao (Southeast University), Geng, Keke (Southeast University), Liu, Zhichao (Southeast University), Wang, Ziwei (Southeast University), Yin, Guodong (Southeast University)

SVM-LO: An Accurate, Robust, Real-Time LIDAR Odometry with Segmentation Voxel Map for Autonomous Vehicles

Scheduled for presentation during the Invited Session "Modeling, Optimization and Game Control of Human-Machine Interaction Behavior in Intelligent Transportation Systems" (ThAT10), Thursday, September 26, 2024, 10:50−11:10, Salon 18

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on January 13, 2025

Abstract

Self-localization is an essential technology for autonomous vehicles. This work proposed a simple and effective LiDAR odometry based on the segmentation voxel map termed SVM-LO. Unlike other LiDAR odometry, the SVM-LO is unlimited by LiDAR scanning modes and environment changes. Our approach consists of two components: the front-end, which acquires key points and the local map, and the back-end, which obtains the robust poses by adaptive mix iterative closest point (AM-ICP). We design an adaptive voxel classifier for extracting more stable key points and constructing the segmentation voxel map. To guarantee system real-time performance, we propose the adaptive dual-subsampling module and optimized the store strategy of the local map. We conduct extensive experiments comparing SVM-LO with state-of-the-art technology on different datasets, the results demonstrate that the SVM-LO is generic, robust, and accurate. Our system is able to run in realtime under multi-scale scenarios, narrow scenarios, and degradation scenarios. SVM-LO eschews reliance on LiDAR scanning characteristics and other sensors, thus it can be widely used in a variety of autonomous driving scenarios.