A SLAM System with Direct Velocity Estimation for Mechanical and Solid-State LiDARs

Abstract

Simultaneous localization and mapping (SLAM) is essential for intelligent robots operating in unknown environments. However, existing algorithms are typically developed for specific types of solid-state LiDARs, leading to weak feature representation abilities for new sensors. Moreover, LiDAR-based SLAM methods are limited by distortions caused by LiDAR ego motion. To address the above issues, this paper presents a versatile and velocity-aware LiDAR-based odometry and mapping (VLOM) system. A spherical projection-based feature extraction module is utilized to process the raw point cloud generated by various LiDARs, hence avoiding the time-consuming adaptation of various irregular scan patterns. The extracted features are grouped into higher-level clusters to filter out smaller objects and reduce false matching during feature association. Furthermore, bundle adjustment is adopted to jointly estimate the poses and velocities for multiple scans, effectively improving the velocity estimation accuracy and compensating for point cloud distortions. Experiments on publicly available datasets demonstrate the superiority of VLOM over other state-of-the-art LiDAR-based SLAM systems in terms of accuracy and robustness. Additionally, the satisfactory performance of VLOM on RS-LiDAR-M1, a newly released solid-state LiDAR, shows its applicability to a wide range of LiDARs.

LMBAO: A Landmark Map for Bundle Adjustment Odometry in LiDAR SLAM

“…Lego-LOAM [8] adds ground segmentation to improve the accuracy of feature extraction of LOAM [7]. VLOM [11] introduces spherical projection [12], [13] to obtain the spatial structure of point clouds, and extracts feature points from spherical images in a novel way, achieving accurate correlation of lines and planes. Some deep-learning based approaches [14], [15] are also used to extract 3D features from sparse point clouds.…”

“…Global Feature Map in Local BA: The use of BA described above is still limited to the mapping step, which requires a separate, previously performed scan-to-scan estimation step as odometry. Instead of performing scan-toscan alignment, VLOM [11] creatively predicts the pose of a new frame immediately from the previous frame with a constant motion model. Feature points extracted within the frame can be directly associated with global features, and used in local BA to optimize velocities and poses.…”

“…This coarse maintenance method without filters results in very limited prior constraints available in the map. For this reason, we improve the map maintenance strategy in [11] to fully exploit the prior constraints and introduce an appropriate marginalization mechanism on the sliding window to reduce the computational cost.…”

“…In fact, the current frame can also improve the estimates of the historical frames, which in turn improves the estimate of the current frame. The introduction of bundle adjustment (BA), which jointly optimizes the pose of multiple frames, avoids error accumulation, and effectively lowers the drift in LiDAR SLAM [9]- [11].…”

“…To provide enough information for optimization, the length of the BA sliding window of map refinement is generally large (set to 20 in [9]). However, when BA is used for odometry, the computational cost is extremely limited due to the high real-time requirements, which means that the length of the sliding window is restricted (set to 8 in [11]). Thus, only information from recent frames is available in the map, and previous constraints from earlier frames cannot be used.…”