Abstract
Detection And Tracking of Moving Objects (DATMO) is essential for mobile mapping systems to generate clean and accurate point cloud maps, as dynamic targets in real-world scenarios often deteriorate system performance. This research presents a robust LiDAR-SLAM system incorporated with a real-time dynamic objects removal module to improve the accuracy of 6-DOF pose estimation and map precision. The proposed method efficiently clusters sparse point clouds of moving objects and tracks them independently to mitigate their influence on odometry and mapping. Furthermore, a probabilistic map fusion method based on free-space theory is employed in the back-end to refine the final point cloud maps. Evaluation on datasets collected from crowded environments demonstrates competitive results against state-of-the-art systems in both pose estimation and mapping.
Introduction
Most Simultaneous Localization and Mapping (SLAM) algorithms operate on the fundamental assumption that the environment is static. However, urban environments are frequently populated with moving objects such as vehicles and pedestrians. When a LiDAR sensor captures these dynamic targets, two primary issues arise: odometry drift, where the system erroneously locks onto moving objects to estimate its trajectory, and map "ghosting," where moving objects leave blurred trails of points in the final 3D reconstruction. This research addresses these challenges by identifying, tracking, and removing dynamic elements in real-time.
Methodology
The proposed system architecture consists of three parallel threads: Laser Mapping, Laser Odometry, and the DATMO module.
Detection and Tracking of Moving Objects (DATMO)
To handle sparse LiDAR data, the system utilizes a multi-stage detection pipeline:
- Ground Filtering & Segmentation: Ground points are removed, and the remaining points are coarsely segmented.
- Convex Hull Filtering: Points are filtered based on proximity to the ground and range to isolate likely dynamic targets.
- Density-Adaptive DBSCAN: Standard clustering often fails as point density decreases with distance. The system employs an adaptive clustering method where the search radius and minimum point count adjust based on sensor characteristics.
- Tracking: Clustered objects are tracked using an Extended Kalman Filter (EKF) and Multiple Hypothesis Tracking (MHT). Confirmed moving objects are stripped from the data before they can interfere with odometry.
Probabilistic Map Fusion
To further refine the map, the system applies Free-Space Theory. By maintaining occupancy probabilities for voxels in an octree structure, the system can statistically "vote out" transient dynamic points missed by the front-end. If a laser beam passes through a voxel to hit a target further away, the probability that the intermediate voxel is occupied decreases.
Experimental Evaluation
The system was evaluated using a Clearpath Jackal UGV equipped with a Velodyne VLP-16 LiDAR. Two datasets were collected at Wuhan University:
- Playground Dataset: A 988m loop in a crowded area with high pedestrian traffic.
- Building Dataset: A 352m loop characterized by complex geometry and foliage.
Performance was benchmarked against industry standards: A-LOAM and LeGO-LOAM.
Results and Discussion
The proposed system demonstrated significant improvements in both trajectory stability and map cleanliness. In the Playground dataset, the system showed substantially less accumulated drift compared to LeGO-LOAM, with discrepancies of only 2 to 5 meters at loop closures where other systems struggled.
Qualitative results indicated that while standard systems left heavy regions of blurred "ghost" points, the proposed system produced clean, static-only maps suitable for high-fidelity modeling. The DATMO thread maintains real-time performance at 10Hz, matching the laser scanner's frequency.
Conclusion
By treating dynamic objects as entities to be tracked rather than simple outliers, the proposed LiDAR-SLAM system achieves high robustness in crowded urban environments. While the system is highly effective, future work will explore integrating semantic information via deep learning to better distinguish between complex dynamic objects and static background features, further reducing instances of over-segmentation.
Summary updated