Smart Traffic Signal Optimization using Reinforcement Learning and Computer Vision

Project Abstract

This study presents a novel framework for dynamic traffic signal control that combines reinforcement learning with computer vision to optimize urban intersection performance under real-world variability. We design a closed-loop system where a deep reinforcement learning agent, informed by a convolutional neural network-based perception module, continuously observes traffic states from live video feeds and sensor data to determine adaptive signal phase and duration. The perception module extracts multi-objective features such as vehicle count, queue length, turning movements, pedestrian presence, and incidents, while ensuring robustness to occlusions, weather conditions, and camera calibration differences through data augmentation and domain adaptation techniques. The reinforcement learning component employs a hybrid action space that integrates traditional phase plans with flexible green time allocation, enabling both fixed and adaptive sequencing. We implement a reward function that balances throughput, average waiting time, queue stability, and fairness across approaches, with priority adjustments for emergency vehicles and high-priority corridors. To ensure safety and regulatory compliance, the system incorporates constraint-based safety checks and seamless fallback to fixed-cycle strategies during perception or communication failures. The methodology is validated through a multi-layer evaluation pipeline (i) simulation-based experiments using a high-fidelity traffic simulator to compare against fixed-time and actuated controls across varying traffic demands, incident scenarios, and network topologies; (ii) field tests on a controlled intersection with calibrated cameras and inductive loop sensors to assess transfer performance and real-time latency; and (iii) ablation studies to quantify the contributions of perception accuracy, reward shaping, and exploration strategies. We analyze key performance indicators including average travel time, intersection throughput, average cycle length, stop frequency, vehicle delay distribution, pedestrian wait times, and energy consumption of the infrastructure. Results demonstrate statistically significant reductions in total delay and queue lengths while maintaining or improving throughput under peak conditions, with faster adaptation to demand fluctuations and incidents compared with traditional optimization schemes. The framework exhibits resilience to partial sensor outages and demonstrates scalable potential for deployment across interconnected corridors with cooperative signaling strategies. We discuss computational efficiency, model generalization across cityscapes, and data privacy considerations related to camera data. The work contributes a transferable end-to-end solution that leverages modern artificial intelligence techniques to enhance traffic efficiency, safety, and sustainability, supported by a reproducible experimental setup, open-source components, and guidelines for urban interoperability and policy integration.

Project Overview

What This Project Is About

---

A practical study that explores how traffic signals can be controlled more intelligently by using computer vision to read traffic flow and reinforcement learning to adjust signal timings over time for smoother traffic movement.

---

---

The Problem It Addresses

---

Many cities suffer from congestion due to fixed traffic signal schedules that don’t adapt to real-time conditions, causing longer wait times and higher emissions. This project seeks a smarter, adaptable approach to reduce delays and improve safety.

---

---

Objectives of the Project

---

1. Understand how traffic data can be collected and analyzed.
2. Implement a basic computer vision module to detect vehicle flow at intersections.
3. Apply reinforcement learning to optimize signal timings based on observed traffic.
4. Evaluate improvements in wait times, throughput, and emissions estimates.
5. Assess the practicality and limitations for real-world deployment.

---

---

What You Will Do Step by Step

---

Step 1: Review related work on smart signals and vision-based traffic sensing. Step 2: Set up a simulated intersection environment. Step 3: Build a simple camera-flow estimator to measure cars waiting and passing. Step 4: Train a reinforcement learning agent to adjust lights. Step 5: Run experiments comparing adaptive vs fixed schedules. Step 6: Analyze results and discuss constraints. Step 7: Document findings and potential real-world considerations.

---

---

---

---

Expected Outcome

Demonstrated improvement in traffic flow metrics (shorter waiting times, higher throughput) using adaptive signals guided by computer vision data, with a clear understanding of limitations and steps toward real-world testing.