Context Aware Crowd Tracking and Anomaly Detection via Deep Learning and Social Force Model

Abstract

The world's expanding populace, the variety of human social factors, and the densely populated environment make humans feel uncertain. Individuals need a safety officer who generally deals with security viewpoints for this frailty. Currently, human monitoring techniques are time-consuming, work concentrated, and incapable. Therefore, autonomous surveillance frameworks are necessary for the modern day since they are able to address these problems. Nevertheless, hardships persist. The central concerns incorporate the detachment of the foreground from the scene and the understanding of the contextual structure of the environment for efficiently identifying unusual objects. In our work, we introduced a novel framework to tackle these difficulties by presenting a semantic segmentation technique for separating a foreground object. In our work, Super-pixels are generated using an improved watershed transform and then a conditional random field is implemented to obtain multi-object segmented frames by performing pixel-level labeling. Next, the Social Force model is introduced to extract the contextual structure of the environment via the fusion of a novel chosen particular histogram of an optical stream and inner force model. After using the computed social force, multi-people tracking is performed via three-dimensional template association using percentile rank and non-maximal suppression. Next, multi-object categorization is performed via deep learning Feature Pyramid Network. Finally, by considering the contextual structure of the environment, Jaccard similarity is utilized to make the decision for abnormality detection and identify the unusual objects from the scene. The invented framework is verified through rigorous investigations, and it obtained multi-people tracking efficiency of 92.2% and 89.1% over the UCSD and CUHK Avenue datasets. However, 95.2% and 93.7% abnormality detection efficiency is accomplished over UCSD and CUHK Avenue datasets, respectively.