Video surveillance based crowd activity analysis

© 2019 Meng YangVideo-based crowd motion analysis is an important problem in surveillance applications. Tasks such as identifying anomalous crowd motion patterns, finding sudden changes in the size of a crowd, event detection, and monitoring in public places have gained considerable attention due to safety and security considerations. While crowd analysis covers a wide range of applications, we focus on anomalous event detection and crowd behaviour identification. Although several approaches have been proposed for crowd scenario analysis in recent years, there still remain limitations in the existing techniques as well as challenges such as handling complex scenarios in the context of real-world scenes. One of the challenges in anomalous event detection is how to define anomalies. To address this challenge, we propose a spatio-temporal feature representation scheme that can be combined with a hyperspherical clustering approach to perform online anomalous crowd analysis in video surveillance. This system can flag abnormal events such as people standing statically or loitering in crowded scenes. Our next challenge is to devise a technique that can perform short-term anomalous object detection, as well as performing long-term anomalous motion/behaviour detection, while guaranteeing high accuracy and efficiency. To achieve this goal, we propose a deep learning architecture that comprises of a stacked denoising auto-encoder (SDAE), deep belief network (DBN) and plane-based one class support vector machine (PSVM) for automatic feature extraction and learning. The DBN-PSVM combination helps achieve better detection accuracy by means of dimensionality reduction to obtain a few high level representative features from the DBN before applying the PSVM. Considering that deep learning based architectures still suffer from high computational complexity, we model the crowd activity as graphs and address the problems of abnormal event detention and crowd behaviour identification in video sequences by using graph mining techniques. We first introduce a scalable clustering-based method that can be used to detect changes in crowd movements, including sudden running, merging, and splitting events. Furthermore, we develop a novel architecture to translate the crowd events in videos into a graph stream analysis problem, where major flow regions are represented as nodes and the crowd density is modelled as the edge weights in the graph. Anomalous crowd behaviour events are identified as the change points in the time series of graphs. Finally, we propose a more general graph-based architecture to achieve both anomalous event detection and crowd behaviour identification, where we take each moving object as a node, and use similarity measures (such as direction, magnitude, velocity and geo-distance) as the edge weights. The video sequence can then be represented as a time-varying graph sequence. We use a max-flow/min-cut approach to model the changes in the network and identify anomalous incidents and objects in the scene. Throughout the thesis, we empirically evaluate our proposed methods on a variety of benchmark and real-life video datasets of crowd movements in public spaces