The team developed an algorithm to detect obstacles in subway tunnels using LiDAR data. The process involves identifying rails, reconstructing the path, and checking for objects within the train's clearance gauge. The method does not rely on pre-built maps or specific obstacle classification models.
The detection process begins by analyzing LiDAR point clouds to locate rail pairs. The algorithm filters points based on geometric constraints, such as the 1520 mm track gauge, and evaluates candidates using density histograms and surface quality metrics. The system also identifies the contact rail to help track the path.
During a test case, the system processed 175,688 points to identify a large square object located 96.96 meters from the sensor. A subsequent measurement package confirmed the detection after 0.30 seconds. The approach uses a clearance gauge 2.1 meters wide and 3 meters high from the rail heads.
To ensure accuracy, the algorithm incorporates historical data from 11 previous LiDAR frames to reconstruct the path, while obstacle confirmation relies on a fresh set of three frames. The system uses Cauchy loss functions during optimization to minimize the impact of outliers on the detection results.
