Adaptive traffic light control using vision-based deep learning for vehicle density estimation
Weerasak Karoon, Peeranut Chuasuai, Pearploy Thipprasert, Nachasa Khongchu, Piyaboon Kunakornjittirak, Thitirat Siriborvornratanakul · 2024
In today's urban landscape, traffic congestion poses a significant and far-reaching problem impacting cities, economic development, and individual well-being. The root of this challenge lies in the ineffective traffic light management system. In this study, we integrated intersection videos captured by cameras into our system, aiming to improve solutions by evaluating two vehicle detection methods: background subtraction (MOG) and YOLOv3. YOLOv3 outperformed MOG in accuracy, leading to its adoption. We employed the DeepSORT algorithm for vehicle tracking and counting, crucial for determining green light duration. Using Arduino, we controlled the green light based on these calculations. Our experiments confirmed YOLOv3's superiority in vehicle detection, while our prototype system demonstrated proficiency in detection, counting, and green light duration calculation. However, room for improvement remains in vehicle type classification.