Proactive Anomaly Detection for Smart Crowd Management Using Hierarchical Temporal Memory

Abstract

Crowd anomalies, including overcrowding or collisions, may endanger humans and property and have occasionally ended in tragedy, serious injuries, fatalities and financial loss. Therefore, crowd management solutions become inevitable for large-scale events as well as daily activities to avoid or mitigate crowd anomalies and ensure people’s safe, smooth movement. Crowd management involves real-time monitoring and modelling of crowds’ temporal patterns to project effective mechanisms that support quick diversion from dangerous situations. Smart Crowd Management (SCM) systems use the Internet of Things (IoT) and Machine Learning (ML) technologies, which leverage the intelligence, speed, and efficiency of the crowd management process. These SCM solutions aim to predict crowd status or detect potential anomalies. However, there are several challenges associated with the existing crowd prediction and anomaly detection solutions, which restrict the capabilities of current SCM solutions. For example, most adopted learning mechanisms cannot efficiently distinguish actual anomalies from the inherent IoT sensory noise. Moreover, current crowd management solutions employ offline learning models that fail to capture the dynamic spatio-temporal relations of the online crowd data streams. In addition, existing solutions overlook crowd sequence learning, which is essential to accommodate the continually evolving noisy sensory inputs to correctly predict sequence of crowd status. Unfortunately, current crowd anomaly detection techniques are reactive and lack the proactive capability to forecast potential anomalies and respond quickly to serious crowd issues. The rapid advancement in deep hierarchical models that can learn from continuous data streams has not been fully investigated in the crowd management context. For example, Hierarchical Temporal Memory (HTM), a biologically-inspired sequence-based memory system, has shown powerful capabilities for application domains that require unsupervised online sequence learning, noise-tolerance and modelling of temporal information. This research aims to investigate the application of HTM for developing a proactive crowd management framework to provide early alerts of potential crowd risks. Due to the lack of representative crowd datasets, the entire flow for generating and preparing synthetic crowd datasets that exhibit multiple examples of abnormal behaviour, such as high density and contra-flow, is presented. SIMulated Crowd Datasets (SIMCD) were developed using the MassMotion simulator. As a primary research, HTM-based crowd prediction and reactive anomaly detection models were examined and evaluated. The F-scores of the proposed reactive HTM-based anomaly detection outperform k-Nearest Neighbour Global Anomaly Score (kNN-GAS), Independent Component Analysis-Local Outlier Probability (ICA-LoOP) and Singular Value Decomposition Influence Outlier (SVD-IO) by more than 18%. For the online crowd severity level prediction, HTM achieved F-scores comparable with Long Short-Term Memory (LSTM) of 91.3% and 90.6%, respectively. However, in offline prediction LSTM slightly outperforms HTM by around 2.6 %. Then, the thesis introduces a novel alertness framework for the proactive detection of crowd anomalies, such as high densities, unstable speed or movement against the expected flow. The alertness framework managed to proactively detect potential crowd anomalies with an accuracy of 91.5%. This proactive capability of the alertness framework would enable the crowd management system to detect potential anomalies early enough to overcome the response time issues seen in reactive anomaly detection approaches. Finally, a new attention-based prediction framework was proposed which mimics the attention mechanism of human beings. This proposed framework is an attempt to enhance the learning capability of the crow