关键词: 视频异常检测, 扩散模型, 不确定性, 记忆池, 生成模型

Abstract: Video anomaly detection is becoming more and more important in modern society. Because there are various fuzzy behaviors in videos and the categories can not be exhaustive, it is difficult to define normal and abnormal based on single-classification methods. In response to the aforementioned issues, a Conditional Diffusion Probabilistic Model for Video Anomaly Detection (CDiffuVAD) is proposed. Firstly, a normal content generator is designed to improve the content accuracy of the image generated by the model. It highlights the distribution pattern of normal samples through a memory pool and leverages the diffusion probability model to learn the complex distribution of video data. Secondly, implicit motion conditions are designed and introduced to learn the spatiotemporal features of video segments. The bidirectional optical flow information is used as the implicit motion condition of the diffusion process, and the coordinate normalization method is used to provide the coordinate embedding of the segment frame, so as to realize the learning and fitting of the motion trend in the multi-frame sequence data and alleviate the sensitivity of the model to the hard normal information in the video. Finally, experiments that the proposed method achieves 85.7%, 75.5% and 65.7% accuracy of frame-level AUC on Avenue dataset, ShanghaiTech dataset and UBnormal dataset respectively, indicating that it can find diverse features in normal samples and the effectivity in video anomaly detection tasks. 

Key words: video anomaly detection, diffusion model, uncertainty, memory pool, generative model