BeautyMap: Binary-Encoded Adaptable Ground Matrix for Dynamic Points Removal in Global Maps

TL;DR

BeautyMap addresses dynamic objects contaminating global LiDAR maps by representing the environment as binary-encoded vertical occupancy matrices and performing fast bitwise comparisons to detect dynamic regions. It introduces an adaptable ground adjustment using MAD-based coarse ground extraction and a fine ground segmentation step, followed by a static restoration module that uses visibility-aware masking and reverse virtual ray casting to protect static points. The approach achieves state-of-the-art dynamic-point removal performance with high efficiency on KITTI and semi-indoor benchmarks, and is open-sourced for community use. This enables robust, high-fidelity global maps for accurate localization and planning in open-world robotics scenarios.

Abstract

Global point clouds that correctly represent the static environment features can facilitate accurate localization and robust path planning. However, dynamic objects introduce undesired ghost tracks that are mixed up with the static environment. Existing dynamic removal methods normally fail to balance the performance in computational efficiency and accuracy. In response, we present BeautyMap to efficiently remove the dynamic points while retaining static features for high-fidelity global maps. Our approach utilizes a binary-encoded matrix to efficiently extract the environment features. With a bit-wise comparison between matrices of each frame and the corresponding map region, we can extract potential dynamic regions. Then we use coarse to fine hierarchical segmentation of the $z$-axis to handle terrain variations. The final static restoration module accounts for the range-visibility of each single scan and protects static points out of sight. Comparative experiments underscore BeautyMap's superior performance in both accuracy and efficiency against other dynamic points removal methods. The code is open-sourced at https://github.com/MKJia/BeautyMap.

Figures (8)

• Figure 1: The results of dynamic points removal in SemanticKITTI dataset, utilizing our proposed approach called BeautyMap. Red: all of the dynamic points (left) and after removed (right). Blue: static points.
• Figure 2: The overview pipeline of our proposed method. The global map is encoded as a matrix, and processed by the adaptable ground adjustment. Scans are similarly encoded after being transformed into the global frame. To realize dynamic region recognition, the query matrix is compared with the corresponding map submatrix bitwise to get a potential dynamic matrix. Finally, incorrectly classified points that are actually static are restored.
• Figure 3: Implementation of our proposed map representation structure. (a) Global cluttered point cloud map. The red refers to dynamic points from ground truth only for visualization and no ground truth label will be used. (b) Map storage structure. Colored voxels mean they are occupied, and free voxels in white. Hierarchical resolutions on z-axis are applied to the ground grid cells shown in brown. (c) The binary-encoded 2D map matrix. The 2D matrix values of the zoom-in region are shown in hexadecimal, with the view of vertical occupancy data on the right. For a better view, each value is logarithmically mapped and visualized in heatmap format.
• Figure 4: Underground outlier points in the black square are apparently shown in the side-view of the global point cloud map. The color bar indicates the point heights in meters.
• Figure 5: Our fine ground segmentation process. The black box zooms in on $C_{sel}$ showing the ground points number ratio segmentation along the black dashed line. Blue points for static and red points for dynamic. The red boxes are potential dynamic grids and the dashed blue boxes with both static and dynamic points are for internal calculation.