Prediction Enhancing Of Video Future Frame For Video Anomaly Detection

The rapid development of artificial intelligence technology has promoted the research of video anomaly detection methods based on deep learning.Video anomaly detection is still a challenging problem due to the ambiguity and diversity of anomalies in different scenes.Recent efforts towards video anomaly detection try to learn a deep autoencoder to describe normal event patterns with small prediction errors.However,there are two main problems in these methods: 1)The existing video prediction models have poor performance in future frame prediction;2)Due to the strong generalization capability of the deep autoencoder,the existing methods can sometimes predict the abnormal inputs well.Thus resulting in abnormal missed detection.Aiming at the problem that the existing video prediction models have poor performance in future frame prediction,a video anomaly detection method based on gradient attention mechanism is proposed.We use appearance constraints based on pixel intensity and gradient and motion constraints based on optical flow in future frame prediction.In addition,we also design a gradient-based attention mechanism to enhance the quality of video future frame prediction.It solves the problem that the video frames generated by most prediction models are unrealistic.The enhanced future frame prediction model provides more real video frames and improves the accuracy of video anomaly detection.Aiming at the problem that the generalization ability of existing deep autoencoders may be too strong,a video anomaly detection method based on prototypeguided discriminative latent embeddings is proposed.We design a prototype-guided memory module to enhance the autoencoder,which can record typical normal patterns in video data.In addition,in order to further improve the way of motion information modeling,we use another branch network to generate RGB difference.The stacked RGB difference contains motion information just like optical flow,and the calculation speed is faster,so our model can learn the temporal regularity.The enhanced future frame prediction model is more inclined to predict normal events than abnormal events and is more sensitive to the occurrence of anomalies.It solves the problem of abnormal missing detection in the method based on deep autoencoders.