A Study on Generalization of Anomaly Detection Based on Spatio-Temporal Graph Normalizing Flows (STG-NF) in Public Spaces

Recently, there has been a growing adoption of deep-learning-based anomaly detection in Closed-Circuit Television systems for enhanced security in public spaces. Most of these video surveillance systems employ an unsupervised learning approach, relying on exclusive training with normal data for anomaly detection. This is achieved through the use of an autoencoder structure that processes RGB data, mainly due to the challenges associated with defining all abnormal behaviors. However, this approach encounters a reduction in recognition accuracy when the environment or location changes. To address this issue, there is a need for an additional step involving the collection of environment-specific data. In this paper, we adopt Spatial-Temporal Convolution Network to analyze skeleton data by constructing a spatiotemporally connected graph from video data and performing graph convolution. The proposed module interprets poses and movements in successive video frames based on information extracted from human joint data. Furthermore, to overcome the limited anomaly detection performance caused by the generalization performance boundary issue in existing autoencoder-based anomaly detection, we propose a normalized flow model to learn normal data and to discriminate anomalous data. Processing skeleton-based data and implementing anomaly detection through normalized flow not only minimizes the impact of environmental factors but also ensures consistent performance across diverse settings. This approach overcomes the limitations associated with autoencoder models relying on RGB data, making it applicable and effective in public spaces. In this paper, we additionally define anomalous behaviors that commonly observed in real public spaces, such as instances of fainting integrated into the dataset for a more practical anomaly detection and robust against environmental factors and it is evaluated using existing open datasets, demonstrating its capacity to operate realistically across diverse environments.