Data Collection and Deep Learning for Traffic and Road Infrastructure Management

Abstract

Managing the traffic of a city or a demographic is a deceptively significant task. The productivity of the people commuting through the region and the pollution caused by their vehicles can be critically impacted by the traffic management of the roads that they are in. Therefore, artificial intelligence systems can be a good alternative to traditional traffic management techniques. A framework for mining real-time data will help build the foundation for intelligent traffic management approaches that can be immediately tested instead of proposing a data collection subroutine every time. In addition to data collection, this paper proposes ways to derive insights, from which both short-term (real-time traffic routing and dynamic traffic signals) and long-term (building flyovers and installing new traffic signals) decisions can be made. Raw data is collected in the form of video feeds from the traffic cameras. Machine learning routines such as Deep Sort and Lucas–Kanade method are used to extract statistics (such as speed of vehicles and vehicle count) from the video feeds. Multi-layer perceptions are used to extract license plate information from the video feeds. This license plate information can be further extended for many security applications.