In recent years, simultaneous localization and mapping (SLAM) based on light detection and ranging has been playing an increasingly important role in computer vision, robotics, and other fields. Existing SLAM frameworks are generally based on the assumption of a static world; however, dynamic objects in the environment, such as walking people and moving vehicles, inherently exist. These objects leave ghosts on the 3D map constructed using SLAM. These ghosts are treated as obstacles in the map, hindering the motion planning of mobile robots, causing errors in LiDAR odometry, which affects the effectiveness of SLAM. Therefore, detection and removal of dynamic points from the point-cloud map are particularly important before performing the corresponding tasks. To address the challenge of mapping in dynamic environments, numerous researchers have proposed methods for building static maps that leverage the geometric discrepancies between individual scans and a map cloud. Despite these advancements, two primary limitations remain in many current systems. First, when the pose estimations derived from scan registration become imprecise, the geometric relationship between the current scan and map cloud is compromised, leading to the erroneous exclusion of numerous static points. Second, most existing approaches overlook instance-level information, which causes points belonging to moving objects to persist within a map cloud. To address these limitations, this study proposes a LiDAR static mapping method based on spatiotemporal constraints.

To enhance the effectiveness of dynamic point removal, this study proposes a new method for building static maps using LiDAR, considering spatiotemporal constraints. First, a ground segmentation algorithm based on candidate height values is introduced to improve ground segmentation. Second, a dynamic point detection algorithm based on grouped optimization and a pseudo-occupied grid is proposed to group scan frames, thereby introducing a new feature descriptor to perform the initial detection of dynamic points within the groups. Third, by leveraging temporal information before and after grouping, a dynamic point region-growing algorithm based on interframe matching distance constraints and a clustering verification strategy based on edge detection are combined to address the issues related to false positives and false negatives in dynamic point detection. Finally, all dynamic points are removed to obtain the ultimate static map.

In this study, sequences 00 (frames 4390?4530), 01 (frames 150?250), 02 (frames 860?950), 05 (frames 2350?2670), and 07 (frames 630?820) were selected as static map construction benchmarks, with the numbers in parentheses indicating the start and end frames. The selected frames encompassed diverse scenes such as rural areas, highways, and intersections, with a substantial presence of dynamic objects for simulation validation. To quantitatively assess the efficacy of our algorithm, we relied on three key metrics: static point-cloud preservation rate (PR), dynamic point-cloud removal rate (RR), and their harmonic mean F1'. The existing open-source algorithms, Removert-RM, Removert-RM+RV, ERASOR, and the proposed method were compared in terms of qualitative and quantitative results. The results indicate that the proposed method outperformed the other methods, excelling in PR and F1' values, by 6.3% and 2.0%, respectively, compared to the erase sample overlap and relatedness (ERASOR) algorithm. A comparison of the running time of the aforementioned methods is provided. To comprehensively evaluate the algorithm, real-world data captured using a VLP-16 LiDAR in avenue and parking garage were used for validation. Removert-RM, ERASOR, and the proposed method were qualitatively and quantitatively compared on these two scenarios. The proposed method achieved dynamic point removal F1' values of 92.65% and 90.03% for the avenue and parking garage, respectively, demonstrating substantial advancements over the classic Removert-RM and ERASOR algorithms.

To address the issue of decreased quality in static maps generated by LiDAR due to the presence of dynamic objects in real-world environments, this paper proposes a spatiotemporally constrained method for constructing static maps using LiDAR. This method leverages the temporal inconsistency of dynamic point clouds and introduces a novel feature-descriptor constraint to obtain preliminary detection results for dynamic points. Subsequently, by utilizing the temporal information before and after multistep checks, growth operations were conducted on the dynamic points. The experimental results consistently demonstrated that compared with existing methods, the proposed approach exhibits superior performance in both the removal of dynamic points and preservation of static points. Future work will involve incorporating point-cloud distribution characteristics, utilizing statistical hypothesis test to eliminate falsely detected points with low confidence, and adapting thresholds based on the motion speed of dynamic objects to further enhance the quality of the static maps generated by LiDAR.