Challenge / Goal
Traffic in the urban district was originally controlled based on static cycle regimes of traffic lights developed by traffic engineers. These regimes changed three times a day at fixed times and did not respond to frequent traffic fluctuations with a rather large variance. This inefficiency of control led to significant drivers' time losses and severe congestions.
Our goal was to implement and deploy a flexible control scheme, based on state of art AI techniques, which allows for real time monitoring of traffic and real time control of traffic lights.
1. We implemented a computer vision module that analyses video streams from the cams installed at intersections, recognises and tracks all vehicles in the visibility region (80-100 meters), and extracts information about their positions and speed in real time.
2. We implemented an infrastructure for training reinforcement learning agents, using a traffic simulator calibrated with real world data obrained from the computer vision module. The infrastructure is general enough to be able to incorporate all available (now or in the future) sources of traffic data and to train an optimal agent recommending phases for traffic lights controller. This agent is a core of our recommendation module.
3. We integrated our computer vision and recommendation modules in the central traffic management system of the transport department of Moscow and provide recommendations for phase switchings in real time.
Citizens can comment about traffic situation at specific intersections, using an online platform of the traffic department.
Want to learn more about the lessons learned, financial details and results?