A Real-Time Proximity Sensing System Using Point Cloud Segmentation in an Octree Structure

This paper presents a novel third-party proximity sensing system for external observation to detect potential collisions between objects in real time. The proposed system, implemented using the Robot Operating System, consists of three modules: 1) Collecting world point clouds using low-cost 3D depth sensors, 2) Segmenting the collected point clouds into target objects and background using an improved min-cut algorithm, and 3) Efficiently detecting potential collisions by converting point cloud data into octree structures and performing radius-based nearest neighbor search. The system's performance is enhanced by introducing filtering techniques to reduce computational burden and improve responsiveness. Experimental results, using a 7-axis robot arm and objects of various shapes, demonstrate the system's ability to accurately separate objects, trigger real-time potential collision warnings, and outperform state-of-the-art methods in terms of processing speed and data handling capacity. The developed system’s modularity and implementation ensure compatibility with different robots and sensors, making it suitable for diverse applications requiring real-time proximity detection in complex environments.