Deep Learning for Traffic Light State Classification

This project develops a robust traffic light classification system for autonomous vehicle applications, comparing two architectural approaches. A baseline was first established using a custom CNN arch, MobileNetV2 and VGG16 trained on augmented full images (200 labelled samples, resized to 224×224). The improved pipeline introduced YOLOv8s with CVAT hand-annotated bounding boxes to first localise the traffic light region, then classify it. This detection-then-classify approach eliminates background noise that caused misclassifications in the baseline. Key Features: - Augmentation strategies included random rotation, zoom, channel shift, Gaussian blur, and simulated rain/fog to mimic real-world driving conditions - CVAT was used for hand annotation of 200 real world images of green and red traffic lights - YOLOv8s demonstrated strong generalisation on unseen real-world images across complex urban scenes