Abstract: Currently, the application of point cloud edge features in coal mines is primarily focused on areas such as identification of comprehensive mining faces and autonomous driving in open-pit mines, with relatively limited application in unmanned tunnel inspection tasks. In the context of belt roadway inspection in coal mines, edge extraction of belt conveyor enables robots to ascertain their spatial pose relative to the conveyor, thereby providing environmental information crucial for executing inspection tasks and assessing coal flow transportation conditions. Introduces a Quadtree-based Points Edge Extraction (QBPEE) technique, which combines projection with a quadtree method. Initially, through an analysis and classification of environ-mental characteristics in underground coal mines under point cloud scenarios, a sensor deployment scheme tailored to belt corridors is proposed, reducing the requirement for lidar units to a single device. Addressing the sparsity issue in belt conveyor point clouds, the method integrates RANSAC plane extraction with the quadtree structure to achieve edge extraction, demonstrating enhanced performance in handling segmenta-tion and clustering issues of belt conveyor point clouds compared to three other typical algorithms. Based on data sampling from the Ningtiaota Coal Mine's belt roadway, the SOR filtering thresholds are set at KD=25 and MulThresh=0.1, with a plane extraction threshold for the conveyor belt at 0.02. Subsequently, rviz and Gazebo are utilized in a joint simulation to convert a densely scanned point cloud model of the belt roadway, obtained via handheld laser scanning, into a tunnel simulation scenario with surfaces. The extraction results are tested under various simulated and real-world operational conditions of the inspection platform, con-sidering factors such as different motion scenarios, radar obstruction rates, and environmental reflections. Over 10,000 consecutive frames, accuracy remains above 96.33% across varying motion conditions, drops to a minimum of 79.23% when radar obstruction is below 30%, and reaches 88% under conditions of full saturation water layer coverage, with a per-frame computation time of 36ms, which is lower than other typical algorithms. This work thus provides a key technological underpinning for unmanned tunnel inspec-tions in underground mines. The constructed roadway simulation data and part of the test video are derived from Github.