A LiDAR returns millions of points; a safety system has to return one bit. The work is not the sensor — it is the pipeline that segments the ground, finds the rails, watches only the track corridor, tells a person from an animal from debris, and raises a train alert inside a response time that actually protects the train.
| Sensor | Strength | Limit |
|---|---|---|
| LiDAR | Direct range and 3-D shape; works in darkness and glare | Cost; heavy rain and snow scatter the return |
| Camera | Rich texture, low cost, strong classification | No direct range; fails in darkness and glare |
| Radar | Long range, weather-robust, direct velocity | Coarse shape; small objects hard to classify |
| Thermal | Sees warm bodies in total darkness | No range; occlusion and sun-warmed clutter |
A LiDAR frame is millions of returns, and almost all of them are the ground, the ballast, the rails and the scenery. The safety question is a single bit — is something in the train’s path — so the pipeline’s first job is subtraction. Ground-plane segmentation levels the scene and removes the track bed, and because a trackside unit vibrates and settles, that calibration has to run continuously rather than once at install.
What is left is still not an alarm. The system detects the rail lines and builds a dynamic area of interest — the track corridor plus a small margin either side — so that a person standing safely beside the line does not trigger the train alert while the same person between the rails does. The area of interest moves as the geometry changes, which is why it is computed live rather than drawn once on a map.
Only inside that corridor do classification and tracking matter. Objects are separated by size and density so a small animal or a scrap of debris is not treated as a person, and each object is tracked for pose and velocity so the system knows whether it is moving onto the track or away from it. The number that decides whether any of this is worth doing is the detection-to-alert latency: an alert that arrives after the braking distance has closed is not a safety system, and a false alarm that stops a train has a real cost of its own, so the threshold is a balance, not a maximum.
Because a camera has no direct range and struggles in darkness, glare and weather — exactly the conditions a trackside safety system has to work in. LiDAR returns range and 3-D shape directly, day or night, which is what defining a track corridor and measuring how far away an object is both need. Cameras and thermal can complement it, but the range and geometry come from LiDAR.
By watching only the track corridor, and by classifying objects on size and density before alerting. A small animal or a piece of debris does not match the profile of a person, and an object safely beside the line is outside the dynamic area of interest. Tracking adds intent — whether the object is moving onto the track — so a safe crossing is not treated as an intrusion.
Detection-to-alert latency. The system is only useful if the alert reaches the train while there is still distance to act, so the response time is designed against the braking distance. It is balanced against the false-alarm rate, because an alert that needlessly stops a train carries its own operational cost.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.