Graph convolutional neural network for skeleton-based video abnormal behavior detection

Abstract

Video anomaly detection aims to detect abnormal events given only normal events where pedestrians regularly walk in surveillance videos. It is popular to leverage encoder-decoder-based reconstruction or prediction methods, to model a normal distribution upon normal data. Whereas, the background noise of reconstructed or predicted results may harm the final performance. To tackle this human-related task, we introduce a spatial temporal graph convolutional networks-based prediction network for skeleton-based video anomaly detection, which detects anomalies based on skeletons, thus, alleviating the noise from complex backgrounds. Specifically, we build a normal graph describing graph connection of joints in normal data. Then, a fully-connected layer is utilized to predict the future joints. Finally, the future joints in normal events can be well predicted while the abnormal ones lead to a large error. To our knowledge, this is the first work to apply graph convolutional networks on skeleton-based video anomaly detection. Experiments show that our proposed normal graph achieves the state-ofthe-art performance, compared to those image-level reconstruction-based methods, image-level prediction-based methods, as well as skeleton-based RNN-based methods.