Using Computer Vision Systems to Train an Autonomous Drone to Follow Railroad Tracks for Inspection

Utilizing unmanned autonomous vehicles (UAVs) can make the inspection of many structures much more efficient. However, when it comes to the inspection of railroad tracks, a problem that arises is the lack of a good GPS signal for the UAV to navigate the tracks for inspection. This is because many railroads pass through rural areas and tunnels that may not have good connection. The goal of this project was to develop a program that allows a UAV to autonomously follow railway tracks without the need for GPS. The approach taken was to use the UAV’s camera to collect frames of video that could be processed to find the location of the rails for the drone to follow. The images would be processed using a trained U-Net convolutional neural network, Canny Edge Detection, a Probabilistic Hough Transform, and a custom pixel-by-pixel rail tracing algorithm. This program was tested by passing in images of railroad tracks taken from drones to simulate frames of a UAV’s live video feed being passed into the program. The current results of this project are that the rails can be correctly identified and highlighted in some images, however the program still needs to be optimized to work on a wider variety of images. In conclusion, this project was able to create a program to recognize railroad tracks using computer vision that can be utilized by a drone to navigate railroads for inspection without the need for GPS, making UAVs much more viable for the inspection of tracks.