Video anomaly detection based on scene classification

As a significant research hotspot in the field of computer vision, video anomaly detection plays an essential role in ensuring public safety. Anomaly detection remains a challenging task given the complex situation in public areas and the large random distribution of crowds. The density of people in the same scene varies greatly due to the instability of the pedestrian volume. Specifically, the characteristics of crowd distribution mainly include low density, small aggregation and dispersion, or large aggregation and severe occlusion. Considering the large difference between high-density and low-density crowd characteristics, we propose an anomaly detection algorithm based on scene classification in order to obtain better anomaly detection result. Firstly, we propose a novel scene classification method, which uses pre-trained YoloV4 model to detect the number of people in the video frames and generate heatmaps, and extracts pixel features through the Double-Canny algorithm to represent the occlusion degree of the crowd. Furthermore, K-Means clustering is used to adaptively divide the scene into sparse and dense. Secondly, the Generative Adversarial Network (GAN) based on prediction and reconstruction is introduced to detect anomalies respectively, and the final accuracy is achieved by combining the detection accuracy of both networks. Finally, experiments on three benchmark datasets demonstrate the competitive performance of our method with the state-of-the-art methods.