Background Subtraction-based Unlearned Hazardous Object Detection Algorithm for Ceiling Environments

We propose a background subtraction-based algorithm that uses images from an RGB-D camera to detect unlearned hazardous objects in ceiling environments, thereby facilitating efficient ceiling exploration. We constructed a dataset by compiling various images of ceiling structures and employed the YOLOv8s-seg model to recognize ceiling components. Recognized ceiling structures and regions exceeding a specified depth threshold were designated as background and removed in the first stage. A U-map based on depth information was then used to detect object regions, which enabled the identification of various types of potentially hazardous objects. The positions of the detected objects were continuously updated using a Kalman filter based tracking model, and their relative positions were visually displayed on a 2D map. To validate the performance of the algorithm, we implemented it in a system and conducted hazardous object detection experiments in a ceiling-simulated testbed and an actual ceiling environment. The findings confirmed the applicability of the algorithm in real-world settings.