Reactive and Proactive Anomaly Detection in Crowd Management Using Hierarchical Temporal Memory

Abstract

An effective crowd management system offers immediate reactive or proactive handling of potential hot spots, including overcrowded situations and suspicious movements, which mitigate or avoids severe incidents and fatalities. The crowd management domain generates spatial and temporal resolution that demands diverse sophisticated mechanisms to measure, extract and process the data to produce a meaningful abstraction. Crowd management includes modelling the movements of a crowd to project effective mechanisms that support quick emersion from a dangerous and fatal situation. Internet of Things (IoT) technologies, machine learning techniques and communication methods can be used to sense the crowd characteristic /density and offer early detection of such events or even better prediction of potential accidents to inform the management authorities. Different machine learning methods have been applied for crowd management; however, the rapid advancement in deep hierarchical models that learns from a continuous stream of data has not been fully investigated in this context. For example, Hierarchical Temporal Memory (HTM) has shown powerful capabilities for application domains that require online learning and modelling temporal information. This paper proposes a new HTM-based framework for anomaly detection in a crowd management system. The proposed framework offers two functions: (1) reactive detection of crowd anomalies and (2) proactive detection of anomalies by predicting potential anomalies before taking place. The empirical evaluation proves that HTM achieved 94.22%, which outperforms k-Nearest Neighbor Global Anomaly Score (kNN-GAS) by 18.12%, Independent Component Analysis-Local Outlier Probability (ICA-LoOP) by 18.17%, and Singular Value Decomposition Influence Outlier (SVD-IO) by 18.12%, in crowd multiple anomaly detection. Moreover, it demonstrates the ability of the proposed alerting framework in predicting potential crowd anomalies. For this purpose, a simulated crowd dataset was created using MassMotion crowd simulation tool.