Computer Vision Project - Traffic Light Detection

For the course “Bildverarbeitung, Maschinelles Lernen und Computer Vision”, I have been part of a project to detect traffic lights in images and videos as well as the state of the traffic light (red, red-yellow, yellow, green, off), to be used in driver-assistance systems and autonomous vehicles. The code is available on GitHub.

We split our pipeline into multiple separate steps, my contribution to the project was the implementation of the traffic light State Detection model using a custom ResNet architecture, implemented in PyTorch. The model was trained on our self-annotated data from the DTLD dataset provided by driveU. The final model achieved an accuracy of 97.7% on the test set.