4-D Radar-Camera Vector Map SLAM With Dynamic Object Removal Mask and Two-Stage Loop Detection

Abstract

A vector map containing lane information is essential to perform global or local path planning for autonomous driving. Vector maps, also termed high-definition (HD) maps, are typically developed using high-cost LiDAR or a combination of cameras and deep learning. In this article, the first complete real-time vector map simultaneous localization and mapping (SLAM) system using the emerging 4-D radar and low-cost cameras is proposed. First, a dynamic object removal mask (DORM)-based visual-4-D radar odometry is proposed, which incorporates a velocity-adaptive and distance-dependent radius function to ensure robust performance in dynamic urban environments. By considering both the object’s absolute velocity and its distance from the sensor, our method effectively suppresses dynamic features while preserving distant static landmarks. The experimental results indicate that the proposed method outperforms state-of-the-art LiDAR SLAM and other techniques in dynamic scenarios. Second, a two-stage loop detection method is suggested using the vector map generated by the inverse perspective mapping (IPM) and the 4-D radar $Z$ projection image. Experimental validation demonstrates that odometry drift is reduced through pose graph optimization-based loop closure. A demonstration video related to this work is available at the following link: https://youtu.be/pPb2ze24F30