CAD-Mesher: A Convenient, Accurate, Dense Mesh-based Mapping Module in SLAM for Dynamic Environments

TL;DR

CAD-Mesher presents a plug-and-play dense mesh mapping module for SLAM that robustly operates in dynamic environments by coupling a two-stage dynamic removal pipeline with Gaussian-process-based incremental meshing and adaptive keyframe processing. The method, which integrates with various LiDAR odometries, filters dynamic points both coarsely via visibility and finely via voxel-based Bayes occupancy, while constructing dense meshes through GP regression and continuity checks. Key contributions include the explicit two-stage dynamic removal, a sliding-window keyframe aggregation with adaptive downsampling, and a point-to-mesh registration scheme that refines pose estimates to improve both meshing quality and localization accuracy. Experiments on five public datasets demonstrate real-time performance (up to 8–10 Hz) and superior meshing and localization accuracy, especially in dynamic and sparse LiDAR scenarios, with publicly accessible code and video.

Abstract

Most LiDAR odometry and SLAM systems construct maps in point clouds, which are discrete and sparse when zoomed in, making them not directly suitable for navigation. Mesh maps represent a dense and continuous map format with low memory consumption, which can approximate complex structures with simple elements, attracting significant attention of researchers in recent years. However, most implementations operate under a static environment assumption. In effect, moving objects cause ghosting, potentially degrading the quality of meshing. To address these issues, we propose a plug-and-play meshing module adapting to dynamic environments, which can easily integrate with various LiDAR odometry to generally improve the pose estimation accuracy of odometry. In our meshing module, a novel two-stage coarse-to-fine dynamic removal method is designed to effectively filter dynamic objects, generating consistent, accurate, and dense mesh maps. To our best know, this is the first mesh construction method with explicit dynamic removal. Additionally, conducive to Gaussian process in mesh construction, sliding window-based keyframe aggregation and adaptive downsampling strategies are used to ensure the uniformity of point cloud. We evaluate the localization and mapping accuracy on five publicly available datasets. Both qualitative and quantitative results demonstrate the superiority of our method compared with the state-of-the-art algorithms. The code and introduction video are publicly available at https://yaepiii.github.io/CAD-Mesher/.

Figures (7)

• Figure 1: The global mesh map constructed by CAD-Mesher on the KITTI07 sequence. The zoomed-in regions (highlighted by red and yellow boxes) demonstrate the good performance of our method on presenting sufficient details.
• Figure 2: The system is divided into four modules. The keyframe processing module is designed for selecting keyframes and performing coarse dynamic removal to ensure registration quality. In the pre-processing module, keyframes within the sliding window are aggregated, followed by adaptive downsampling to obtain more uniform point clouds. The meshing module conducts continuity test to exclude outliers and leverage GP for mesh construction. Finally, the optimization module performs point-to-mesh registration and fine dynamic removal using voxel-based probabilistic method, obtaining the refined pose estimation and generating a continuous, accurate, and static global mesh map.
• Figure 3: Coarse dynamic removal. The top row displays the range image at frame $k$. The middle row shows the range image of the aggregated point cloud at time $k-1$. The bottom row illustrates the absolute difference image between the two. The color of each pixel represents the distance value. The points in the absolute difference image is over the threshold $r_{th}$ are removed, especially for dynamic points highlighted by yellow box and white circle.
• Figure 4: The effect of our dynamic removal approach. The blue dashed box indicates dynamic ghosting in the environments. (a) Using only visibility-based dynamic removal. (b) Using only voxel-based probabilistic dynamic removal. (c) our two-stage coarse-to-fine dynamic removal.
• Figure 5: Qualitative results. The top row highlights the effect of our dynamic removal, with the yellow box indicating dynamic ghosting. The bottom row shows the robust performance of our method with sparse-channel LiDAR, where green boxes indicate the density or inconsistencies of meshing.