A Framework for Quick and Accurate Access of Interesting Visual Events in Surveillance Videos

Abstract

AbstractIt is a challenge to quickly and accurately access interesting visual events in explosively increasing surveillance video data. In this paper, we develop a novel video summarization framework to accurately detect and quickly browse interesting visual events in large-scale surveillance videos. The method firstly detects interesting foreground objects by running a set of pre-trained object classifiers on foreground object candidates which are generated by a background subtraction algorithm. Then a Hungarian algorithm based multi-objects tracking program is followed to obtain accurate and complete motion trajectories of detected foreground objects. Finally, each interesting visual event is compactly represented by a synthetical snapshot, which makes it convenient to quickly access interesting visual events in long videos. Experiments on challenging surveillance videos show our framework outperforms existing video summarization systems in the detection accuracy of interesting visual events.KeywordsSurveillance VideoForeground ObjectVideo SummarizationForeground RegionHungarian AlgorithmThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.