Visual analysis of dense crowds / Kok Ven Jyn

The steady worldwide population growth with continuing urbanization renders the formation of crowd by chance a norm. The mere existence of crowd has the prospect of progressing into a hazardous scene. Consequently, visual analysis of dense crowds is a growing research topic in the domain of computer vision. Conventional visual analysis methods are mostly object-centric, thus, are neither suitable nor capable of analyzing dense crowd. Hence, this thesis proposes novel solutions to analyze images and videos of dense crowds, which contain hundreds to thousands of individuals. The main objective are, first, to obviate the difficulty of segregating individuals in dense crowd scenes to infer dense crowd segments, secondly to estimate the number of individuals and finally to detect unusual events, by exploiting spatial and temporal cues readily available from the scenes. Dense crowd segmentation generally serves as one of the essential steps for further visual analysis of the dense crowds. The thesis first demonstrates the significance of simplifying dense crowd scenes into structurally meaningful atomic regions for dense crowd segmentation. This proposed approach is formulated using the concept and principles of granular computing. It shows that by exploiting the correlation among pixel granules, structurally similar pixels can be aggregated into meaningful atomic structure granules. This is useful in outlining natural boundaries between crowd and background (i.e. non-crowd) regions necessary for dense crowd segmentation. Moreover, the proposed approach is scene-independent; thus it can be applied effectively to dense crowd scenes with a variety of physical layout and crowdedness. Second, this thesis presents an approach to utilize irregular patches conforming to the natural outline between crowd and background to estimate the number of individuals in dense crowd scenes. As opposed to most of the existing approaches that uses pixel-grid representation, the proposed density estimation approach allows a model to adapt itself to the arbitrary distribution of crowd where the underlying spatial information of scenes can be accurately extracted. Here, a direct mapping is established between the extracted features and the number of people. Third, to detect saliency in dense crowd scenes, low-level features extracted from the crowd motion field are transformed into a global similarity structure. This global similarity structure representation allows the discovery of the intrinsic manifold of the motion dynamics, which could not be captured by the low-level representation. Most importantly, unlike conventional methods, the proposed approach does not require tracking, and prior information or model learning to identify interesting / salient regions in the dense crowd scenes. These proposed approaches are validated by using public dataset of dense crowd scenes. From the empirical results, it is anticipated that the collective analysis of this thesis will constitute a complete dense crowd analysis system that is able to infer regions of dense crowds, estimate crowd density and identify saliency in mass gathering for proactive crowd management