Video Anomaly Detection Framework Based on Motion Consistency

Most methods rely on unsupervised learning due to the limited availability of anomaly data. However, most of the current unsupervised learning methods are based on deep self-encoders, which do not pay enough attention to the consistency of the motion process. Therefore, we propose a video anomaly detection framework based on motion consistency (VADMC).The framework uses a CVAE network as a generator to generate predicted frames. In order to increase the reconstruction error of the CVAE network, we embed a memory module in the optical flow coding features, which is used to memorize the feature distribution of the normal patterns. The reconstruction error is increased by perturbing the a priori distribution, thus increasing the reconstruction error. A discriminator is used to discriminate the generated optical flow maps to ensure the consistency of the forward and backward motions of the normal samples. We conducted experiments on three public datasets to demonstrate the effectiveness of the VADMC framework. The accuracy on the UCSD PED2, CHUK Avenue, and Shanghai Tech datasets reached 97.2%, 76.3%, and 76.2%, respectively. Compared with previous state-of-the-art methods, our method shows competitive results.