Computer vision for autonomous trains: obstruction detection

Case study

Innovation in autonomous transport is booming but has yet to reach broad application due to challenges related reliability and safety. To assist on this front, Thales Canada collaborated with Zetane to develop computer vision models that promote the safety of autonomous trains. Our aim was to produce an object detection algorithm capable of identifying obstructions on train tracks, like a sleeping cow or a boulder; identifying an obstruction should signal the train to stop.

Here we share a summary of this project, displaying components of our dataset and the training of our model. Central to our development process was us conducting explainable AI (xAI) assessments to avoid developing unreliable and potentially inaccurate “black-box” algorithms. In such a “mission-critical” context, we need to prove that our automated computer vision capacities were indeed identifying obstructions on train tracks and avoid assessing irrelevant details in images. Following a Grad-CAM assessment (being a popular xAI Python library), we noted initial iterations of our models were interpreting the dark entry to a tunnel as an object. Subsequent debugging tasks removed this error, where visualizing this problem in Zetane helped streamlined our development process.

To conclude our project, we produced a simulation using LiDAR data to test if our object-detection models would signal a mock-up of a train to stop once it detected a boulder on the tracks.