Adaptive Traffic Signal Control
Traditional traffic lights operate on fixed timers, which is inefficient. This project uses computer vision to count vehicles and Reinforcement Learning (RL) to decide which lane gets the green light and for how long, adapting to real-time surges in traffic.
Technology Stack
System Architecture
Computer Vision
OpenCV and YOLOv8 for vehicle detection and counting.
RL Engine
Deep Q-Learning (DQN) agent trained in a simulated environment.
Simulation
SUMO (Simulation of Urban MObility) for testing and validation.
The Challenges
Training the RL agent to handle edge cases like emergency vehicles.
Coordinating multiple intersections to prevent bottleneck downstream.
Processing high-resolution video feeds at low latency.
The Solutions
Integrated a priority-based reward system in the RL algorithm for emergency vehicle detection.
Implemented a multi-agent RL approach where neighboring signals communicate their flow states.
Used TensorRT for hardware-accelerated inference on the edge.
Key Results & Metrics
Real-time traffic analysis
RL-based optimization
Reduced wait times