Robust real-time traffic monitoring using corners motion statistics

In this paper we describe a virtual sensor to monitor traffic using videos taken from street poles located at junctions in cities. The sensor measures vehicle counts, and the individual speed of each car. The sensor relies upon a layer that detects feature points (corners) and estimates their motion. The computational cost of adding other video-based virtual sensors is negligible The system has been evaluated on extensive recordings (24/7) in real scenarios. These videos and the ground truth measurements have been made public for future evaluations. The overall vehicle count error is about 5%, the average speed error is about ±5 Km/h. The proposed sensor can be used both for automatic traffic regulation and for law enforcement.