AI-Powered Traffic Signal Optimization using Computer Vision and Predictive Modeling

The Problem

Cities currently control traffic lights reactively. Signals change after congestion has already formed. This reactive approach causes excessive idling, fuel waste, massive emission spikes, and unnecessary acceleration cycles.

Furthermore, traditional traffic models rely heavily on average vehicle speed. This fundamentally misses the real stop-and-go behavior at intersections, which is scientifically the largest contributing factor to pollution in urban environments.

Architecture Approach

System Pipeline:

I engineered this solution as a complete, end-to-end pipeline processing real-world data feeds:

• Vehicle Detection: Deployed YOLOv8 to detect and classify different vehicle types directly from existing traffic video feeds, ensuring high accuracy frame-by-frame.
• Vehicle Tracking: Applied the DeepSORT tracking algorithm to follow each unique vehicle across frames. This step was critical to calculating accurate velocity and acceleration data for individual cars rather than generalizations.
• Emission Estimation: Instead of relying on a flawed generic average speed calculation, I implemented a physics-informed Vehicle Specific Power (VSP) model. This allowed the system to compute the real engine load and distinct emission operating modes for each vehicle.
• Predictive Modeling: Trained a multivariate LSTM (Long Short-Term Memory) neural network on the historical state data to predict near-future congestion patterns and upcoming emission spikes before they happen.
• Signal Optimization: Finally, an optimization module ingests the LSTM predictions and dynamically alters traffic signal timing. The objective function minimizes carbon output and reduces high-acceleration events rather than explicitly just minimizing wait time.

Results & Impact

By simulating the optimized signal timings against the baseline reactive control logic, the system demonstrated significant improvements in maintaining flow and reducing the severe stop-and-go emissions typical of congested intersections.

This project proves that combining modern computer vision (CNNs), time-series forecasting (RNNs/LSTMs), and applied physics modeling with real-world infrastructure data is a viable path toward intelligent, green cities.