The registration of traffic data of public roads is not only a compulsory task of local governments, but the traffic count provides valuable information of various kinds. It can be used to estimate the pollution emission and load of a road, which contributes to the operation of the current facilities and to the appropriate planning of new investments. Moreover, real-time traffic monitoring systems are able to detect and even to avoid traffic jams, too.
Nowadays, manual surveillance are more and more being replaced by automatized methods. Besides the radar-based systems and inductive loop detectors installed in the road, the video-based traffic detection is among the most popular methods. However, its real-time implementation, by means of image processing, is still challenging in resource-limited embedded systems. The popular algorithms of background subtraction and feature point tracking are usually associated with a computational cost being too high to meet the accuracy and speed requirements.
In this work, I introduce a system for traffic flow estimation, based on the processing of a video stream of a camera installed above the road. My goal was to devise an intelligent sensor that can be installed within a street-lighting lamp, which operates in real-time on the given resource-limited hardware. It is able to determine the number and type of the vehicles passing by with an acceptable accuracy, and also to estimate their speed and following distance.
The operation and architecture of the core software and the video processing task is presented in detail in this work. The main idea is the combination of various efficient video processing methods: background subtraction, tripwire-based counting, data mining and classification algorithms of low computational cost, respectively. The benefits of each algorithm is exploited via their appropriately combined use. The supplementary software and hardware elements, which facilitate the remote calibration, testing, and the configuration process are discussed. Moreover the accuracy and speed testing process and the results are presented along with possible ways of further improvement.