Smart social junction traffic control using reinforcement learning on real data

To alleviate congestion, specialized lanes for public transportation and carpooling, are set up as Fast Lanes (FL). However, these solutions often lack adaptability, resulting in either overloaded or underutilized lanes. Barzilai et al. [5], proposed a flexible FL managed by traffic volume and social priority, which is a preference criterion based on driver characteristics or travel purpose. They utilized Reinforcement Learning (RL) algorithm, in a basic junction setup using artificial random data. The current study extends this work, to a more practical and effective solution by simulating an actual complex junction with real traffic data based on vehicle detection from surveillance cameras using YOLO and TrafficDataLandBox tools. High accuracy levels, between 89% and 98.7%, were achieved for the vehicle detection task depending on the tool and daytime. To prove the effectiveness of the social priority criterion, we arbitrarily chose a lane assumed to be with high social priority and set it as a FL. We then implemented an extended Q-Learning algorithm for Traffic Signal Control (TSC) according to the traffic and the FL. The results showed that there was indeed an improvement in the average time of the vehicles crossing the junction on the FL without decreasing the crossing time of the other vehicles by much. This study is a step forward in a wide TSC system for several junctions that will consider the traffic flow on nearby junctions, while enabling dynamic FL that can change over time.