Detecting Anomalous Human Trajectories in Indoor Spaces

Abstract

Anomalies in indoor human trajectories often relate to urgent situations (e.g., violent attacks, terrorism, and fire). Detecting anomalous human trajectories can improve safety and instantly handle risks in working spaces. Therefore, this thesis studies methods for anomaly detection in indoor human trajectories. In particular, we first study abnormal indoor human trajectory types, which may occur in real-world environments. Next, a framework for detecting anomaly trajectories based on the density method is proposed. In addition, this thesis develops a trajectory clustering-based anomaly detection framework. Finally, we extend this work by studying a deep learning model to learn trajectory characteristics for detecting anomalies. First, anomaly types of indoor human trajectories are studied. In particular, anomalous trajectories are essential for evaluating anomaly detection methods. However, since anomalous trajectories may not be available in datasets, they are needed to create for evaluation. In the literature, there are often two methods: giving a hypothesis for anomalies, and syn- thesizing anomalies and injecting them into datasets. In the first method, anomalies are assigned based on the difference in occurrence frequency and movement behaviour between trajectory groups in datasets. In the second, anomaly trajectory types are generated and injected into the datasets for evaluation. Secondly, a density method-based framework is proposed for detecting anomalous trajectories. A trajectory is normal if it is near to other trajectories in datasets. In other words, the number of its neighbour trajectories is high. Thus, this framework detects anomalous trajectories based on their densities in datasets. In particular, two trajectories are neighbour to each other if their distance is smaller than a distance threshold. In detecting anomalies, a trajectory is detected as an anomaly if its density is smaller than a density threshold. We propose a new method for determining distance and density thresholds. The next chapter proposes a framework based on a clustering algorithm to detect abnormal indoor trajectories. In this framework, the abnormality of a trajectory is determined based on the relationship between it and clusters in datasets. First, a distance metric is proposed to improve the accuracy of distance metrics for indoor human trajectories. Then, the Epsilon parameter of Density-Based Spatial Clustering of Application with Noise (DB- SCAN) is determined using a new DCVI metric. To find normal trajectory clusters in the dataset, DBSCAN is used. In detecting anomalous trajectory, if a trajectory does not belong to any clusters, it is marked as an anomaly. To take advantage of the power of neural networks in anomaly detection tasks, a Transformer encoder and self-organizing map-based deep learning model called TENSO is proposed to learn trajectory characteristics. In particular, to learn internal characteristics of normal trajectories, the Transformer encoder with the self-attention mechanism is used. In addition, to learn clusters of normal trajectory representations in latent space, the self-organizing map (SOM) is used. In the training phase, the TENSO model is trained using a total loss of trajectory reconstruction and SOM losses. In the anomaly detection phase, a test trajectory is evaluated to determine whether it is an anomaly based on trajectory reconstruction errors and the quantization error on the SOM. Finally, proposed frameworks are evaluated using two real trajectory datasets: MIT Badge and sCREEN. The results show that the proposed frameworks detect trajectory anomalies effectively, especially the TENSO model-based framework.