Visual Object Removal in LiDAR Point Clouds for Dynamic Environment Representation

Abstract

Industrial environments often contain dynamic or temporary objects, such as mobile containers, trash bins, workers, and forklifts, which may degrade LiDAR-based maps by introducing transient structures into the reconstructed representation of the environment. This work investigates the removal of objects of interest from LiDAR point clouds in an indoor industrial scenario using modality-specific instance segmentation and LiDAR-camera projection. For this purpose, a synchronized, calibrated, and annotated multi-sensor dataset was collected, combining RGB images with LiDAR point clouds, as well as LiDAR-derived range and reflectance representations projected onto the image plane. The dataset contains 5016 synchronized frames, corresponding to approximately 8.5 minutes of acquisition at 10 Hz, matching the LiDAR operating frequency. From this data, three modality-specific segmentation datasets were generated: RGB, depth-map, and reflectance-map. YOLO26s-seg instance segmentation models were trained independently for each modality to estimate binary masks of the objects of interest. These masks are then transferred to the 3D domain through LiDAR-camera projection, allowing the corresponding points to be removed from the original point cloud. Beyond offline point cloud filtering, the filtered data are used to generate maps with an in-house SLAM/Localization backend, allowing an analysis of how each modality contributes individually to the cleanliness and consistency of the final map. The segmentation results indicate strong performance for RGB modality, while depth-map and reflectance-map representations achieve lower but still useful performance. SLAM and Localization performance show improvements when dynamic elements are removed from the point cloud. These findings highlight an important trade-off between visually rich sensing modalities, which provide detailed cues for segmentation, and LiDAR-derived image representations, which may enable lower-cost or single-sensor perception pipelines for dynamic-object removal in industrial mapping.