Computational techniques improve vehicle tracking

Professor Nikolaos Papanikolopoulos (pictured right) of the University of Minnesota’s Department of Computer Science and Engineering in the USA has introduced a traffic video event detection system that uses a novel technique to ‘learn’ to recognize events that fit a profile established by the operator.

Developed with graduate student Harini Veeraraghavan, the system represents traffic events as strings of actions, for example, when a vehicle moves through certain areas of the video frame at certain speeds, the computer ‘learns’ what action the movement represents.

Computers do not have a human’s natural capacity to recognize patterns in events, so the researchers use a semi-supervised learning method to teach the computer to recognize events in video data. Instead of programming in every possible variant of each action, the computer system is provided with a small number of clear examples, which helps the computer’s pattern recognition system recognize related strings of actions and improve its accuracy as it processes more data.

In trials, raw video data from a four-way arterial intersection was fed into a conventional video-based data collection system. The researchers claim that the system performed well compared to machine-vision systems based on other event-detection techniques, even though the data included uncontrolled vehicle movements and different lighting conditions.

The research was sponsored by the Intelligent Transportation Systems Institute.
 

26 November 2008