Identifying human activities in megastores through postural data to monitor shoplifting events

Abstract

In recent years, modeling activity patterns for understanding events and human behavior has drawn prominent attention in research. Multiple methods have been proposed for developing automated vision systems that are capable of inferring accurate semantics from the moving dynamics. The multi-disciplinary nature of Human Activity Recognition (HAR) methods and the expanding technologies in this field inspire continual updates in existing methods. However, a cost-effective solution is still needed to recognize human activities like shoplifting in an occluded environment. With this motivation, we present a novel approach to identify human stealing actions by analyzing the postural information of the human body. This approach involves extracting 2D postural body joints of a human being from the captured frame. Pose encoding and postural feature generation in parameter space are the foremost contributions of this work, which can handle the occluded actions too. The feature reduction is done to scale the features into a smaller dimension with an objective of the computationally efficient and real-time solution. Activity classification is done on the reduced feature sets to detect human shoplifting actions in real-time scenarios. Experiments are performed on the synthesized shoplifting dataset, where the results derived are found more promising compared to other state-of-the-art methods, with an accuracy of 96.87%. Additionally, this method exhibits commendable real-time performance in processing actual store camera footage.