Automatic analysis of crowd dynamics using computer vision and machine learning approaches

University of Technology Sydney. Faculty of Engineering and Information Technology.As the population of the world increases, urbanization generates crowding situations which pose challenges to public safety, security and to the management of crowd. Manual analysis of crowded situations is a tedious job and usually prone to errors. Computer scientists have been trying to bring about the cognitive video understanding abilities of human brains onto video systems using computer vision techniques. However, crowd analysis is a complex phenomenon in computer vision because of the intrinsic difficulties such as perspective distortion, occlusion, shadowing effect and variations in scale, pose and illumination. These difficulties degrade the performance of a computer vision system. This research proposes solutions based on classical machine learning approaches as well as state-of-the-art deep learning approaches to automate the process of visual surveillance. This thesis will investigate three research areas of visual analysis of crowds: crowd counting and density estimation in low-to-medium density crowds, crowd motion analysis, and describing crowd motion with useful motion information. To this end, we investigated the texture features for crowd counting and density estimation. We outline an SVM-based framework for the evaluation of different texture features and also studied the performance of the various combination of features for crowd counting and density estimation. Currently, object detection using Deep Convolutional Neural Networks have shown tremendous improvement over hand-crafted feature extraction techniques. Current state-of-the-art object detectors were used for head detection task. Furthermore, Faster R-CNN was used for counting by detection, and the performance is improved with the proposed Motion Guided Filter by using spatio-temporal information. In this study, we also propose a novel technique of crowd analysis, the aim of which is to detect different dominant motion patterns in real-time videos. It is very important for a computer vision system to capture movement as accurately as possible. For such a system, a motion field is generated by computing the dense optical flow. The performance of our approach is evaluated on different real-time videos. Furthermore, a framework is developed that extract motion information from the crowd scene by generating trajectories using particle advection approach. The trajectories obtained are clustered using unsupervised hierarchical clustering algorithm into flows. The motion information extracted includes, the speed, density and mean direction of dominant flows/global flows. This study will provide real-time and practical solutions for the autonomous vision system used in the surveillance. The analysis of information before the occurrence of any disaster or disruptive event will result in the indication of dangerous and disruptive behaviour which seeks urgent attention