Adaptive Traffic Optimization

Ashish Waghode — Sun, 12 Jul 2026 08:53:20 +0000

Problem Statement: Standard fixed-time traffic controllers operate on rigid, pre-programmed cycles without real-time situational awareness, leading to inefficient green-time allocation and increased congestion at unbalanced intersections.

Project Objective: This project develops an adaptive traffic signal control system that utilizes real-time vehicle counts per approach. By leveraging a dynamic agent [or reinforcement learning agent, if applicable], the system optimizes phase sequencing to maximize throughput and minimize cumulative commuter delay.

Solution:I developed a closed-loop, adaptive traffic management system that replaces static signal cycles with real-time, data-driven decisions through four key components:

Real-Time Perception: Deployed a YOLOv8 and OpenCV pipeline across four directional video feeds to detect, classify, and count queued vehicles within defined regions of interest.

Intelligent Decision-Making: Implemented and compared two Reinforcement Learning agents a baseline Tabular Q-Learning model and a PyTorch based Deep Q-Network (DQN) trained on a custom reward function designed to minimize cumulative intersection wait times.

Visual Validation: Built a custom Pygame environment to simulate intersection physics and signal states, enabling real-time visual auditing of the RL agents' behaviors.

Centralized Control Plane: Wrapped the architecture in a FastAPI web application featuring asynchronous MJPEG streams of the detection feeds, live telemetry endpoints, and a frontend dashboard to monitor vehicle metrics and training performance graphs.

Future Improvements:Support more intersection types generalize the fixed 4-direction setup (N/S/E/W) into a configurable N-way system, so the same agent can handle 3-way junctions, 6-way intersections, or roundabouts just by changing a config instead of the code.

Phase-based signals instead of picking one direction at a time, group compatible directions into phases (e.g., opposing traffic together) like real signal controllers do.

Multi-intersection coordination extend from a single junction to a network of connected signals (multi-agent RL) for corridor-level "green wave" optimization.

Richer inputs factor in queue trends over time, pedestrian/cyclist detection, time-of-day patterns, and emergency vehicle preemption.

Better benchmarking compare fixed-timer vs. Q-learning vs. DQN on wait time and throughput to actually quantify the improvement.

Realistic simulation swap the pygame simulator for SUMO to test against real world traffic patterns before live deployment.

This is a proof of concept, so there's plenty of room to build on whether it's extending the RL agent, adding support for more intersection types, or improving the simulation. Check out the repo, open an issue, or send a PR: Traffic-Optimize on GitHub
(https://github.com/ashishbot120/Traffic-Optimize).

DEV Community: Ashish Waghode

Adaptive Traffic Optimization