An Efficient LiDAR SLAM With Angle-Based Feature Extraction and Voxel-Based Fixed-Lag Smoothing

Light detection and ranging (LiDAR) simultaneous localization and mapping (SLAM), which provides accurate pose estimation and map construction, has a broad range of applications in autonomous driving. However, in some complex scenarios, such as degraded environments, SLAM performance is not up to the requirements of an autonomous driving system. To aim at the abovementioned problems, we propose an accurate and robust LiDAR SLAM method. First, in order to avoid the uncertainty of LiDAR viewpoint variation, an angle-based feature extraction method is proposed based on the equiangular distribution property of the point cloud. Second, in view of the fact that odometry errors accumulate, we constructed a fixed-lag smoothing (FLS) to jointly optimize the poses of multiple keyframes. In addition, to improve the environmental representation of point cloud maps within a sliding window, we maintain two types of abstract voxel maps. Finally, a voxel-based feature matching method based on voxel geometry constraints is proposed for the refinement of the pose transformation within a sliding window. The performance and efficiency of the proposed method are evaluated on the public KITTI, Mulran, and The Newer College dataset benchmarks and the dataset collected by our sensor system. The experimental results show that accurate feature extraction, efficient voxel feature matching, and consistent FLS help our LiDAR SLAM method achieve better performance in multiple spatially and temporally large-scale scenarios compared to other existing state-of-the-art methods.