A survey on deep learning-based real-time crowd anomaly detection for secure distributed video surveillance

Fast and automated recognizing of abnormal behaviors in crowded scenes is significantly effective in increasing public security. The traditional procedure of recognizing abnormalities in the Web of Thing (WoT) platform comprises monitoring the activities and describing the crowd properties such as density, trajectory, and motion pattern from the visual frames. Accordingly, incorporating real-time security monitoring based on the WoT platform and machine learning algorithms would significantly enhance the influential detection of abnormal behaviors in the crowds. This paper addresses various automatic and real-time surveillance methods for abnormal event detection to recognize the dynamic crowd behavior in security applications. The critical aspect of security and protection of public places is that we cannot manually monitor the unpredictable and complex crowded environments. The abnormal behavior algorithms have attempted to improve efficiency, robustness against pixel occlusion, generalizability, computational complexity, and execution time. Similar to the state-of-the-art abnormal behavior detection of crowded scenes, we broadly classified methods into different categories such as tracking, classification based on handcrafted extracted features, classification based on deep learning, and hybrid approaches. Hybrid and deep learning methods have been found to have more satisfactory results in the classification stage. A set of video frames called Motion Emotion Dataset (MED) is employed in this study to examine the various conditions governing these methods. Incorporating an appropriate real-time approach with considering WoT platform can facilitate the analysis of crowd and individuals’ behavior for security screening of abnormal events. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2021.