Feature Extraction of Horizontal Plane and Optimization of 3-D LiDAR SLAM in Indoor Environments

Abstract

Simultaneous localization and mapping (SLAM) is a critical technology in robotics, with LiDAR-based SLAM has shown remarkable success in outdoor environments. However, real-time, robust, and precise state estimation in indoor environments remains a major challenge. This article presents an innovative 3-D LiDAR SLAM framework that incorporates multilayer horizontal plane optimization to address these challenges. Two key innovations distinguish this method. First, the method introduces a plane clustering segmentation (PCS) technique, which segments the raw point cloud based on horizontal and vertical curvatures. This allows for the simultaneous recognition of feature points and various inclined planes. Combined with feedback-based odometry information, this technique enables the extraction of horizontal plane points from tilted LiDAR data, followed by fitting and LiDAR tilt correction. This effectively mitigates issues arising from LiDAR motion and large-angle rotations, which often lead to failure in ground point extraction. Second, the framework integrates edge and surface feature scan-to-scan matching with horizontal plane optimization to refine LiDAR odometry. In the backend, edge, surface, and horizontal plane features are incorporated into both local maps and global horizontal plane constraints, improving pose and map accuracy, particularly in environments with varying ground heights. Compared to existing methods, this approach significantly suppresses pose drift in long-range and multifloor indoor environments. Evaluations on both public and custom datasets demonstrate an average 35% improvement in localization accuracy over state-of-the-art techniques. Overall, this method provides a promising approach for real-time, robust, and accurate state estimation and map building in indoor environments, offering noticeable improvements for indoor LiDAR SLAM applications.