LIE Home · Issue Contents 

Vision-Assisted LiDAR Dynamic Object Filtering for Robust SLAM in Dynamic Environments
Bo Lu, Fangdi Jiang, Ruibing Yang, Shifeng Wang, Aakash Kumar, Cui Wang, Maofeng Liao and Shigeng Song

With the rapid development of autonomous driving and robotics, LiDAR has become a crucial sensor for high precision environmental perception. However, the presence of dynamic objects, such as pedestrians and vehicles, significantly disrupts LiDAR point cloud processing and SLAM mapping, causing degraded localization accuracy and mapping quality. This work proposes a vision-assisted LiDAR dynamic object filtering method that integrates synchronized camera and LiDAR data. The method utilizes the YOLO deep learning model for dynamic object detection and introduces a bounding box expansion strategy to fully cover three-dimensional dynamic targets. To further improve filtering precision and mitigate erroneous removal of static points within bounding boxes, a depth clustering module is employed for fine-grained spatial analysis of point clouds. The proposed coarse-to-fine two-stage filtering process effectively reduces the interference of dynamic objects while preserving critical static environmental information, thereby improving SLAM localization stability and mapping accuracy. Extensive experimental results conducted in both indoor and outdoor scenarios demonstrate the robustness and superior performance of the proposed method in complex dynamic environments.

Keywords: LiDAR SLAM, dynamic object filtering, vision-assisted, autonomous driving, robot localization

Full Text (IP)