A Novel Topology Metric for Indoor Point Cloud SLAM Based on Plane Detection Optimization

Accurate self-localization and navigation in complex indoor environments are essential functions for the intelligent robots. However, the existing SLAM algorithms rely heavily on differential GPS or additional measuring devices (such as expensive laser tracker), which not only increase research costs but also limit the deployment of algorithms in specific scenarios. In recent years, reference-free pose estimation methods based on the topological structure of point cloud maps have gained popularity, especially in indoor artificial scenes where rich planar information is available. Some existing algorithms suffer from inaccuracies in spatial point cloud plane segmentation and normal estimation, leading to the introduction of evaluation errors. This paper introduces the optimization of plane segmentation results by incorporating deep learning-based point cloud semantic segmentation and proposes measurement indicators based on the Plane Normals Entropy (PNE) and Co-Plane Variance (CPV) to estimate the rotation and translation components of SLAM poses. Furthermore, we introduce a ternary correlation measure to analyze the relationship between noise, relative pose estimation, and the two proposed measures, building upon the conventional binary correlation measure. Our proposed PNE and CPV metrics were quantitatively evaluated on two different scenarios of LiDAR point cloud data in Gazebo simulator, and the results demonstrate that these metrics exhibit superior binary and triple correlation and computational efficiency, making them a promising solution for accurate self-localization and navigation in complex indoor environments.