Universal LiDAR Odometry and Mapping With Dual Channel Descriptor

Accurate localization and mapping play a critical role in intelligent vehicles for advanced driver assistance systems and autonomous driving. For the rapidly developing field of autonomy, autonomous systems require an accurate LiDAR SLAM system to adapt to different specifications of LiDARs in various complex scenarios. Therefore, this paper proposes a highly accurate and universal LiDAR SLAM method, which is innovative in the design of LiDAR odometry (LO) and loop detection descriptors. The odometry is designed in two key ways. Firstly, this paper focuses on the 3D spatial information of the point cloud and unifies its representation through voxel gridding, enhancing adaptability to different LiDARs and reducing computation time for feature extraction. Secondly, leveraging the “near dense, far sparse” characteristic of LiDAR, a dynamic registration function is designed to improve the stability and accuracy of the odometry. In addressing loop detection, we have designed a novel dual-channel correlation global descriptor, integrating point cloud height and height difference characteristics. This innovative descriptor reduces the sensitivity to translation and improves the precision and recall of loop detection. We have thoroughly tested our method using various public datasets with different LiDARs in a variety of scenarios and compared it with other well-known methods. Experimental results show that our SLAM technology improves stability and accuracy while maintaining universal compatibility. On the KITTI benchmark, our method ranks among the highest in terms of accuracy and real-time performance, achieving a relative average translation error (RATE) of 0.59 and an average time of 40 ms per scan.