Abstract: The belt conveyor is one of the inspection targets of the inspection robot in the unstructured belt roadway of underground coal mines. Extracting its edges allows the robot to obtain its spatial pose relative to the inspection target, providing environmental information to support the execution of inspection tasks. Currently, most underground edge extraction techniques are vision-based, which struggle to overcome challenges such as low illumination, heavy dust, and dense fog. To address this issue, an explosion-proof 16-line LiDAR was used as the sensor for the inspection robot to acquire the roadway point cloud, reducing the environmental impact on the extraction results. The raw sparse point cloud was preprocessed using statistical outlier removal and passthrough filtering to eliminate noise and irrelevant points. The belt conveyor's point cloud plane was segmented using the Random Sample Consensus (RANSAC) algorithm, and the edge point cloud of the belt conveyor was extracted using a projection-quad tree method. The combined rviz and Gazebo simulation results showed that, under different operating conditions of the robot, the accuracy of belt conveyor edge extraction was no less than 96.33%. When the LiDAR shielding rate was below 30%, the accuracy was no less than 79.23%. Laboratory tests showed that, even when the surface of the belt conveyor had a 100% water distribution and saturated thickness, the edge extraction accuracy was no less than 88%. Overall, this method outperforms the latitude and longitude extremum search method, the curvature threshold method based on KDTree/OcTree, and the adjacent point angle threshold method based on KDTree/OcTree, with an average computation time of only 36 ms, meeting the real-time inspection needs of underground environments.