A Prediction and Knowledge-Enhanced Two-Stream Framework for Anomaly Detection

Video anomaly detection involves identifying unusual behaviors or events in surveillance footage, which is critical for security purposes. However, current methods face challenges due to the unpredictability and context-dependent nature of anomalies. To address these limitations, this paper presents a two-stream anomaly detection framework that combines a prediction stream with a knowledge-enhanced stream. The prediction stream focuses on detecting short-term irregular movements by generating future frames, utilizing attention encoders, temporal shifts, and Swin Transformer modules to better capture temporal patterns in the video. Meanwhile, the knowledge-enhanced stream builds a knowledge base of normal events, helping the model distinguish anomalies more effectively. Experiments on the ShanghaiTech and CUHK Avenue datasets show that the our approach produces highly competitive anomaly detection performances.