Traffic Surveillance in Real-time Using Hidden Markov Models

This paper describes the development of a video-based system for traffic surveillance that works in real-time on a standard PC-platform. The main task of the system, which was developed for and in co-operation with Axicon, is to count passing vehicles to estimate traffic density. More functionality can also be developed on top of this to derive other traffic parameters. The methods use low-level features from predefined regions in each frame to do a preclassification of each region as occupied or unoccupied. Observations from several frames are then combined using a hidden Markov model to eliminate the effect of spurious erroneous detector outputs and to perform the final classification deciding whether a car is passing. The system was tested over a longer period with changing light and weather conditions, and the vehicle counts have been compared with the counts from traditional loop detectors, showing very good results.