Video Anomaly Detection Based on a Motion Pseudo-Sample Synthesizer

Video anomaly detection is a challenging one-class classification problem, as anomalies are rare and diverse in real-world videos. A common approach is to use auto encoders (AEs) to model normal video patterns and detect anomalies based on reconstruction errors. However, this approach may detect anomalies well. In this paper, we propose a novel frame-work for video anomaly detection based on AEs using only motion features. Our framework consists of three components: a gradient-based motion loss function that measures the motion discrepancy between the input and the reconstructed videos, a pseudo-sample synthesizer that generates various abnormal motion patterns to augment the training data and improve the detection performance, and a joint loss function that optimizes the normal motion loss, the anomaly motion loss and the classification loss simultaneously to enhance the discriminative ability of the AE. We conduct extensive experiments on three public datasets and demonstrate that our framework can effectively and efficiently detect abnormal events in complex scenarios.