Motion pattern analysis for far-field vehicle surveillance

Abstract

The main goal of this thesis is to analyze the motion patterns in far-field vehicle tracking data collected by multiple, stationary non-overlapping cameras. The specific focus is to fully recover the camera's network topology, which means the graph structure relating cameras and typical transitions time between cameras, then based on the recovered topology, to learn the traffic patterns(i.e. source/sink, transition probability, etc.), and finally be able to detect unusual events. I will present a weighted statistical method to learn the environment's topology. First, an appearance model is constructed by the combination of normalized color and overall model size to measure the appearance similarity of moving objects across non-overlapping views. Then based on the similarity in appearance, weighted votes are used to learn the temporally correlating information. By exploiting the statistical spatio-temporal information weighted by the similarity in an object's appearance, this method can automatically learn the possible links between the disjoint views and recover the topology of the network. After the network topology has been recovered, we then gather statistics about motion patterns in this distributed camera setting. And finally, we explore the problem of how to detect unusual tracks using the information we have inferred.