Skip to content
ALL WORK
Undergraduate thesis2022AI & MLResearch

Traffic Light Detection with YOLOv3 / v5 / v7

Benchmarking three YOLO generations for real-time traffic light detection under simulated adverse weather.

My undergraduate thesis at RUET. Traffic light detection is a small-object problem at speed — the target is a few dozen pixels, the latency budget is tight, and the failure mode is not academic.

I trained YOLOv3, YOLOv5 and YOLOv7 on the Bosch Traffic Light Dataset augmented to simulate rain, fog and low light, then benchmarked all three on both detection accuracy and inference speed. YOLOv7 reached the highest mAP at 98.3% while also being the fastest of the three — the accuracy/speed tradeoff people usually assume did not hold across generations.

Highlights

  • 98.3% mAP with YOLOv7, the best accuracy and the best inference speed in the comparison.
  • Augmented Bosch dataset simulating diverse weather conditions.
  • Direct accuracy-versus-latency benchmark across three YOLO generations.

Gallery

Keep reading