DIPLOMARBEIT Visual Traffic Surveillance Using Real-time Tracking

This diploma thesis presents an unusual event recognition approach in the field of traffic surveillance. Such events are unusual traffic behaviour like traffic jams, accidents or ghost drivers. An interest-point based tracking algorithm (KLT-tracker) is discussed which pursues features on vehicles through a static camera scene. Tracking data can be collected by observing normal traffic. Then, this data is used to learn a spatio-temporal model of normal traffic behaviour. Thereby, training samples are generated in a learning space by the tracking data. Thus, the spherical probability density function (p.d.f.) of the space can be estimated. We use a Growing Neural Gas in combination with a MDL-based pruning algorithm for unsupervised learning. The former method belongs to the class of soft-competitive algorithms which overcome the problems of ”stranded” reference vectors. In contrast to other works, the number of reference vectors has not to be constant. The algorithm finds an optimal codebook according to the MDL principle. As the p.d.f. only describes points and not trajectories of normal traffic behaviour, behaviour classes of normal traffic have to be learnt additionally. This work presents a novel approach by using the topology of the learning space which is created by Competitive Hebbian Learning. Beside the necessity of recognizing unusual events, it can also be used to analyze the behaviour of drivers at traffic sites like intersections or road works.