Video Anomaly Detection Based on Attention Mechanism

Camera surveillance is widely used in residential areas, highways, schools and other public places. The monitoring scanning of sudden abnormal events depend on humans. Human anomaly not only consumes a lot manpower time but also has large error detection. Video detection based AE (Auto-Encoder) currently the dominant research approach. model highly symmetrical network structure encoding decoding stages. trained by learning standard video sequences, anomalous are later determined terms reconstruction prediction error. However, case limited computing power, complex will greatly reduce efficiency, unnecessary background information seriously affect accuracy model. This paper uses loaded with dynamic prototype units as basic We introduce an attention mechanism to improve feature representation ability Deep separable convolution operation can effectively number parameters complexity. Finally, we conducted experiments three publicly available datasets real scenarios (UCSD Ped1, UCSD Ped2 CUHK Avenue). experimental results show that compared baseline model, our improved 1.9%, 1.4% 6.6%, respectively, across datasets. Compared many popular models, validity verified.