Crowd Abnormal Behavior Detection Based On Sparse Coding

With the worldwide population explosion and city development, there are moreand more densely populated public places. While these places provide peopleconvenience, they also provide places for crime. Therefore, public security is anurgent issue faced by the entire world. As a main method to control the crime inpublic paces, video surveillance is applied more and more widely. Yet, because ofhuman cannot focus on every single details in the whole scene and they may be tiredduring the work, using technological means to automatically analyze the video anddiscover anomalies is a popular research topic in computer vision community.For abnormal behavior detection in crowded scenes, many models andalgorithms have been published at top conference and journals in recent years.Moreover, several publicly available dataset referred to crowds behavior arepublished, too. Although there have been a plenty of works, many flaws still exist inthem. Especially when in extremely crowded areas or occlusion frequently happen,the performance of detection has a lot of space to improve.In this paper, we present a novel method based on sparse coded motion saliencyfor detecting abnormal events in crowded scenes. Unlike existing sparse codingbased approaches, our model dose not need to learn a dictionary. Instead, it directlysparsely represents the motion feathers of the object region using its surroundingareas. The sparse representation error is used to measure the motion saliencyintensity.Additionally, to reflect the situations in the real world more precisely, weassume that: the area with more intense or more disorder motion has moreprobability to exist abnormal behavior. Based on the assumption above, we devisetwo attributes: motion intension attribute and motion consistency attribute anddesign two algorithms to measure these two attributes.To evaluate our approach, two publicly available datasets–UMN dataset andUCSD ped1dataset are utilized to evaluate our approach in detecting globalabnormal event and local abnormal event, respectively. We present the result in theshape of ROC curve. What is more, we compare our experimental result with otherstate-of-the-art approaches published in recent years. The results show that ourmethod achieves the promising performance.