Video anomaly detection based on deep learning

Video anomaly detection is a crucial topic in the field of deep learning. Typically, manually monitoring abnormal behavior in surveillance videos is a labor-intensive task. As society advances, the significance of video anomaly detection continues to grow. Inspired by reconstruction methods and frame extraction for constructing pseudo anomalies in anomaly detection, we propose an autoencoder-based anomaly detection model. In this model, normal samples are represented as average images, and anomaly regions are highlighted by whitening. The model offers several advantages: 1) It significantly reduces the construction of anomalies in normal parts compared to previous reconstruction methods, as there is no need to reconstruct foreground content. 2) Anomaly regions are clearly highlighted by whitening. 3) The model exhibits high sensitivity to individuals and objects with abnormal speeds in the scene, enabling easy identification. 4) Overlaying normal and pseudo-anomalous samples effectively addresses issues related to occlusion of normal parts and insufficient emphasis on anomalous regions. We conducted tests on the USCD Ped2 anomaly detection dataset, achieving a micro AUC of 99.6%. Our model demonstrates competitiveness compared to state-of-the-art models.