Along with the widespread growth of surveillance cameras, computer vision algorithms have played a fundamental role in analyzing the large amount of videos. However, most of the current approaches in automatic video surveillance assume that the observed scene is not crowded, and is composed of easily perceptible constituents. These approaches are not extendable to more challenging videos of highly crowded scenes such as in religious festivals, marathons, sport events, public demonstrations, subways etc., where detecting and tracking individuals is a very difficult task. Therefore, a number of studies have recently begun to focus on the analysis of high-density scenes.
Crowd analysis has recently emerged as an increasingly important and dedicated problem for crowd monitoring and management in the visual surveillance community. Specifically, the estimation of crowd density is receiving a lot of attention and it is of significant interest for crowd safety in order to prevent potentially dangerous situations. In this thesis, our first objective is to address the problems of crowd density estimation (such as people counting, crowd level estimation and crowd motion segmentation) and the second objective is to investigate the usefulness of such estimation as additional information to other video surveillance applications. Towards the first goal, we focus on the problems related to the estimation and the characterization of the crowd density using low level features in order to avert typical problems in detection of high density crowd, such as dynamic occlusions and clutter. We demonstrate in this dissertation, that the proposed approaches perform better than the baseline methods, either for counting the number of people in crowds, or alternatively for estimating the crowd level. Afterwards, we propose a novel approach for crowd density measure, in which local information at the pixel level substitutes the overall crowd level or person count. Our approach is based on modeling time-varying dynamics of the crowd density using sparse feature tracks as observations of a probabilistic crowd function.
The second goal of this study is to explore an emerging and promising field of research in crowd analysis which consists of using crowd density as additional information to complement other tasks related to video surveillance in crowded scenes. First, since conventional detection and tracking methods are not scalable to crowds, we use the proposed crowd density measure which conveys rich information about the local distributions of persons in the scene to improve human detection and tracking in videos of high density crowds. Second, we investigate the concept of crowd context-aware privacy protection by adjusting the obfuscation level according to the crowd density. Finally, we employ additional information about the local density together with regular motion patterns as crowd attributes for high level applications such as crowd change detection and event recognition.