Training Rails
Real time railroad track defect detection
Challenges We Ran Into
Gathering the data. There weren’t any usable datasets for analyzing train track defects so we had to go outside and manually walk across and video Corvallis train tracks. We then annotated the dataset ourselves and trained the model successfully.
Training model
Making the model work with different distances and angles from the train tracks.
We had to retrain the model weights.
Smallefcts Used multi scale segmentation different angles, positions, and environments
Accomplishments That We Are Proud Of
What are you most excited about?
The frontend Google Maps API 95% computer vision model success rate Our live demo showing the model working on mock train tracks.
What We Learned
What did your team learn while building this?
Segmentation parameters Integrating an edge device with a full-stack application Bridging together multiple pieces of hardware (camera, GPS, GPU).
What Is Next
We will improve the frontend design and scale the backend for better performance. We also plan to connect with rail safety organizations like the Federal Railroad Administration and the Association of American Railroads.
Additionally, we will add a frontend chatbot that connects to an MCP server on the Jetson to display system metrics such as CPU usage, GPU usage, and FPS in real time.