Video abnormal behaviour detection based on pseudo‐3D encoder and multi‐cascade memory mechanism

Frame prediction methods based on Auto-Encoder (AE) composed of convolutional neural networks (CNN) are very popular in detecting abnormal behaviour. The predict normal behaviour accurately and incorrectly, which is considered a criterion for abnormality discrimination. However, the emergence problems such as too strong AE representation leading to detection failure, insufficient ability network extract spatio-temporal information, large number model parameters slow running speed leads need method be further improved. In this work, authors propose framework video pseudo-3D encoder multi-cascade memory mechanism (MMP3D). First all, consisting convolution used information from video. Then, (MM) multi-headed prototype attention store aggregate features behaviour, solves some extent problem failure caused by power. Finally, decoder designed 2D deconvolution layers recover information. efficiency superiority our validated Ped2 dataset, Avenue ShanghaiTech dataset.