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 consecutive and dense 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 be easily integrated 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. 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 six 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.github.io/.
As a mapping module in SLAM, the system receives the raw point in the LiDAR coordinate system at current time, and the pose transformation from the LiDAR coordinate system to the global coordinate system estimated by the odometryas the system input. The keyframe, dropped by the proposed adaptive selection mechanism, is added to the database after visibility-based coarse dynamic removal. The keyframes within the sliding window are subsequently aggregated and converted to the world coordinate system W, then uniformly sampled by the adaptive downsampling strategy to enhance system efficiency. Continuity test is utilized to remove outliers and noise. The remaining points are divided into voxels, GP-based meshing is then conducted. In the optimization component, the pose estimated by the odometry is used as a prior for point-to-mesh registration, aligning the current scan to the global map and outputting the finer pose. Finally, after fine dynamic removal using the voxel-based probabilistic method, the current mesh is fused into the global mesh map for publication.
We select several typical sequences from the KITTI, UrbanLoco, and GroundRobot datasets for localization evaluation. Our CAD-Mesher achieves superior overall performance with 1.38% average translation error and 0.50deg/100m rotation error, which outperforms those point cloud, surfel, mesh-based, and deep learning-based methods. Notice that although PIN-SLAM achieves performance second only to ours, it requires significant GPU resources, a characteristic also shared by SuMa and TRLO.
Our algorithm achieves robust pose estimation through the proposed efficient dynamic filtering strategy. Additionally, our method handles dynamic situations more effectively compared with the dynamic LO TRLO.
For the GroundRobot dataset, the usage of sparse channel LiDAR can easily cause significant z-axis drift while presenting a challenge for meshing. Also, our method maintains meshing density, thus extracting more accurate mesh normals, which contributes to point-to-mesh registration, ultimately achieving satisfactory performance.
First, we conduct an evaluation on KITTI 00~10 sequences using the absolute trajectory error (ATE) and relative trajectory error (RTE) as metrics, with the results presented in Table. III. In general, when the LiDAR odometry is integrated with our module, both ATE and RTE improve. An interesting observation is that the accuracy of the combination primarily decreases in the terms of RTE, while APE improves. We analyze that our registration and mesh fusion strategy enable the current scan to revisit the previous region, functioning as an implicit closed-loop, which may result in a slight decrease in RTE and an improvement in global accuracy. Overall, this integration significantly improves the accuracy of the LiDAR odometry across 11 sequences.
Furthermore, we use ATE as metric on the UrbanLoco and GroundRobot datasets to analyze the effect of integrating our module for improving localization accuracy. Compared to the KITTI dataset, the improvement in localization accuracy from this integration is more pronounced on these two datasets. Our dynamic removal strategy significantly enhances the naive odometry in high-dynamic urban sequences. Notably, A-LOAM+Ours prevents the degradation of original A-LOAM in UrbanLoco05. We attribute this improvement to our two-stage coarse-to-fine dynamic removal approach. In sparse-channel sequences, the dense and accurate mesh representation also contributes to suppress drift. In summary, our CAD-Mesher mapping module can seamlessly integrate with various LiDAR odometry systems to further improve localization accuracy. Additionally, the integrated system can effectively cope with highly dynamic scenes and sparse-channel LiDAR data.
In KITTI05, KITTI07, UrbanLoco05, moving vehicles at intersections leave obvious ghosting in the maps for the other mesh baselines, which impedes following navigation applications. both SLAMesh and ImMesh exhibit noticeable artifacts in their maps due to the absence of a dynamic filtering strategy. SHINE-Mapping and PIN-SLAM also contribute to the dynamic removal by incorporating specifically designed loss items during the training process, while a slight ground ripple effect also ensues. In contrast, our method effectively eliminates dynamic components through the proposed two-stage coarse-to-fine dynamic removal approach while maintaining ground flat. In addition, SHINE-Mapping and PIN-SLAM exhibit significant map misalignment due to cumulative errors in the front-end odometry, while VDBFusion shows severe ground holes, which demonstrates that meshing quality is highly dependent on localization accuracy. As a plug-and-play back-end module, our method achieves more accurate and consistent static maps by refining the pose estimation of the odometry.
In the Newer College dataset, low-speed pedestrians leave long ghosting on the mesh maps generated by SLAMesh and ImMesh. VDBFusion, SHINE-Mapping, and PIN-SLAM build relatively accurate maps with filtering dynamic pedestrians, but the ground is missing to varying extents, affecting the completeness of the maps. Our approach also works well for low-speed pedestrians while preserving the integrity of the ground.
In the GroundRobot01 sequence, the sparse-channel LiDAR presents a challenge to registration accuracy and meshing quality. Due to the sparsity of the point cloud, both VDBFusion and SLAMesh produce many holes in the ground, compromising the continuity of mesh map. Although SHINEMapping mitigates the impact of sparsity and generates a dense map, it exhibits stratification in the green box due to odometry drift. Although SHINE-Mapping and ImMesh mitigate the impact of sparsity and generate a dense map, it exhibits stratification in the purple box due to odometry drift. PIN-SLAM is less affected, but the dividing lines of ground stratification remain visible. However, our method provides more accurate poses for meshing by refining pose estimation of odometry, thereby ensuring consistency of meshing.
Our method achieves the highest reconstruction precision in both datasets, although its recall are not as high as that of SHINE-Mapping. We attribute this discrepancy to our coarse dynamic removal method inadvertently removing some static points, and the proposed consistency test inadvertently filtering out slender poles along with noise. Nevertheless, our method achieves the highest F1-Score, demonstrating its overall effectiveness. It is also worth noting that SHINE-Mapping is a deep learning-based approach for offline post-processing, which requires extensive training and cannot operate in real time. Note that our method can run without GPU resources in real-time, while SHINE-Mapping is a deep learning-based approach offline post-processing, and PIN-SLAM \cite{pin-slam} requires extensive GPU consumption.
As shown in Table. VI, our method outperforms all offline, online, and mesh-based methods across all sequences, achieving excellent associated accuracy. On the KITTI00 and KITTI05 sequences, our method is second only to deep learning-based methods. However, DeFlow suffers from the so-called domain adaptation problem, which is evident in its sharp accuracy reduce when tested on the Argoverse2 dataset after being trained on the KITTI dataset. Also, the deep learning-based methods, including SHINE-Mapping and PIN-SLAM, require extensive GPU consumption. In contrast, our method achieves balanced performance across all test sequences in real-time using only a CPU. Additionally, we find that our method performs slightly worse in the terms of SA, which is attributed to misdeletion of some static points during the rough dynamic removal stage.
According to (a)-(d), as the size of the sliding window increases, both the completion and F1-Score of meshing improve, which demonstrates the superiority of our keyframe and sliding window method. However, if the sliding window size is excessively increased, although meshing quality slightly improves, the computational time rises significantly. We consider the improvement is that marginal compared to the significant increase in runtime. Additionally, localization accuracy tends to decrease due to potential interference from the increased noise.
Comparing (c), (e), and (f), the employments of continuity test and dynamic removal modules enhance the accuracy of localization. Although this may lead to a slight decrease in meshing quality due to occasionally inadvertent deletions of static points, we devote to implement a convenient, accurate, and dense mesh-based mapping module in SLAM, which needs to balance localization accuracy and meshing quality, different from an offline mesh generator.
According to (c), (g), and Fig. 7, employing the adaptive downsampling strategy reduces the system time cost by nearly half while maintaining the meshing quality. Also, more accurate mesh maps further enhance point-to-mesh registration accuracy, resulting in a lower ATE. This highlights the advantage of our adaptive downsampling mechanism in terms of time efficiency.
Finally, by comparing (c) and (h), when adaptive keyframe selection is disabled to include all scans, higher recall is achieved due to improved completion of small areas. However, lower precision is observed, likely due to the simultaneous introduction of noise. As discussed in Sec. III-B, the introduction of the adaptive keyframe selection strategy enables more rational input, resulting in a higher overall F1-score, while also reducing system computation time. Additionally, this strategy also improves global localization consistency.
CAD-Mesher