Intelligent traffic light control system based on computer vision

DOI: 10.31673/2412-9070.2026.318120

Abstract

The article presents the development and experimental verification of a prototype of an inteligent traffic light control system based on computer vision and deep learning. The inefficiency of tradi-tional fixed-cycle traffic lights causes chronic congestion, leads to losses of time and fuel, and hinders the passage of priority vehicles (ambulances, fire trucks, emergency services). The proposed system addresses this problem through adaptive phase control based on real-time video stream analysis.
The architecture is implemented in Python using Ultralytics YOLOv8n for object detection, OpenCV for video capture and visualization, FastAPI and Uvicorn for an asynchronous REST API with automatic documentation. The spatial analysis module relies on polygonal zones ("north–south", "east–west", pedestrian), whose configuration is stored in JSON format according to the actual geometry of the intersection. The control logic is implemented as a finite state machine with six phases: green and yellow for the "north–south" direction, red for all directions, green and yellow for the "east–west" direction, red for all directions. The decision module combines time constraints (mini-mum and maximum green phase duration) with dynamic comparison of directional loads and ensures priority passage for emergency vehicles with immediate phase switching. All decisions are recorded with an explanation of the reason for switching. The thread-safe architecture synchronizes access to shared data between the video processing loop and the web server. The REST API provides endpoints for system health monitoring, retrieval of metrics (vehicle count, pedestrians, priority transport, FPS) and manual phase control. An analysis of related research on YOLO, LLM agents and reinforcement learning was conducted, confirming the alignment of the proposed solutions with global trends. Prototype limitations are identified (single intersection, no object tracking, controller emulator) and a development roadmap is outlined: integration with industrial controllers, object tracking, multi-camera support, metrics database storage and redundancy in case of AI system failure. The prototype can serve as a foundation for industrial traffic flow management systems in smart cities.

Keywords: AI, adaptive traffic light control, computer vision, deep learning, object detection, information system, management, video analytics.