Maintenance of train track equipment is cumbersome and inefficient. Plans showing the location of equipment are always out of date and the information is usually only available in PDF form. Because of this, technicians have a hard time locating the equipment that needs to be repaired.

What it does

In our solution, a camera is installed in the driver's cabin on a train. We built a neural network that analyzes the camera feed to tag all equipment on the captured route and correlates the tags with equipment lists. The merged data is then displayed on a Google Map so the technician can easily find the equipment their looking for and plan the quickest route to get there.

How we built it

We used the YOLOv3 recurrent convolutional neural network built with DarkNet and trained it remotely on a NVIDIA GeForce GTX 1080 GPU with a training set of 1000 labeled images in approximately two hours. The EXIF data from the matched images provided us with GPS coordinates for the equipment. The existing list of signals and balises was only tagged with the distance from the origin station, so we downloaded all waypoints of the train track in question from OpenStreetMap and matched them with the GPS coordinates from our network to merge the two datasets. The resulting data was visualized on Google Maps by utilizing the JavaScript SDK.

Challenges we ran into

Getting the YOLOv3 network to run on our remote server required resolving many dependencies. Choosing the correct training configuration required many iterations.

Accomplishments that we're proud of

The performance of the recurrent convolutional neural network, given the small training set and short training time, is surprisingly good. We've managed to match all equipment with objects found in the image dataset.

What we learned

We've learned how to quickly choose the right neural network and then build, train and deploy it in a project.

What's next for YOLOTrain

To further assist technicians in the field, we would want to develop a mobile augmented reality app. When arriving to a track on foot, the app will show the technician on a camera overlay the direction and distance of the closest signals and balises.

Built With

Share this project: