Urban traffic congestion increases travel delay, fuel consumption, and air pollution, especially at intersections controlled by fixed-time signals. Traditional signal plans are usually designed for average traffic demand and cannot respond effectively to sudden changes caused by peak hours, road incidents, public events, or weather disruption. This paper proposes an edge-based adaptive traffic signal control framework that combines computer vision, queue estimation, and reinforcement learning. Roadside edge units process camera streams locally to estimate vehicle count, queue length, and lane occupancy without transmitting raw video continuously. A reinforcement-learning policy then selects signal phases that reduce delay while maintaining pedestrian safety and emergency-vehicle priority. Simulated evaluation shows a 24.6% reduction in average vehicle delay, an 18.9% reduction in queue length, and lower communication overhead compared with cloud-only traffic analytics