Map-based Multi-Object Dynamic Scene Description and Atypical Event Detection

Abstract

The thesis is about the automatic detection of atypical changes and events in dynamic multi-object scenes. Examples for such scenes are streets, parks, train stations, public places, airports, sports fields, highways, shopping malls etc. The actors or objects traveling through such scenes are pedestrians, bicycles, cars, boats, airplanes, sportsmen or customers. Information on the observable states of the actors or objects is made available by sensors. Based on object state vectors, a new approach is proposed for a map-based representation of the scene dynamics. Two methods are introduced, map-based change detection (MBCD) for the detection of overall scene changes and map-based trajectory analysis (MBTA) for the evaluation of individual object trajectories. Both methods work with real-life multi-object scenes without signatures or training. They were implemented with three different scenes and each scene was annotated by ten human participants in a separate experiment. The evaluation of the results shows the strengths and limits of the new approach and suggestions are made where to continue with further work.