RGB-Assisted Depth Completion for LiDAR Occlusion in Manufacturing Environments

초록

Light Detection and Ranging (LiDAR) sensors are widely used in industrial human-robot collaboration (HRC) for real-time spatial awareness and safety control. However, their performance degrades when sensing objects with challenging material properties-such as glossy, dark, or reflective surfaces-leading to occluded regions and missing depth data. These failures pose serious risks in safety-critical environments. This study proposes a lightweight framework that compensates for LiDAR occlusion using monocular depth estimation from RGB imagery. Designed to augment low-cost LiDAR systems without hardware modification, the method restores missing depth by fusing RGB-derived predictions into occluded regions of the LiDAR scan. Evaluation is conducted using a dual-LiDAR setup, where high-resolution depth serves as ground truth. Experiments with the Rainbow Robotics RB5 demonstrate substantial error reduction in occluded zones, validating the framework's effectiveness under industrial conditions. The proposed approach offers a practical solution for enhancing depth perception in manufacturing environments, and lays the groundwork for future integration with 3D point cloud restoration, safety zoning, and real-time robotic control.

키워드