Scout Robotics
Scout mounts sensors on regular trains to inspect the track as they run. I worked on perception: segmenting rails, working out where the train is, and estimating its speed.
- I fine-tuned SegFormer to segment rails and cleaned up broken masks with geometric post-processing. That made switch detection reliable enough to feed a Viterbi pass, which works out exactly which route the train took on the network map.
- To get speed from a single camera, I warped frames to a top-down view and tracked rail ties between frames, using segmentation, Hungarian matching and a Kalman filter. A LiDAR-calibrated depth map gives the true metric scale.
- I moved ROS bag extraction onto a Jetson AGX Thor, using NVIDIA’s GPU JPEG libraries and parallel buffered reads. It ran about 6x faster and saved over $15k in cloud compute.
- I built a detector for fasteners, bolts and clips from a DINOv2 encoder and a DETR decoder. It reaches over 90% precision.
- I managed deployment of the pipelines on AWS for the MTA pilot, and built visualizations of the results for MTA executives.