Video Anomaly Event Detection Methods Based On Deep Learning

Anomaly detection is an active area of computer vision and is widely applied in diverse scenarios,especially in video surveillance.Currently,intelligent video surveillance systems can detect abnormal behavior in real-time and reduce the loss caused by abnormal events.However,it is time-consuming and laborious to obtain and label anomalies in some cases.In this regard,anomaly detection methods based on unsupervised or semi-supervised learning are advantageous.Many existing anomaly detection methods employ feature differences between the input and output spaces to detect anomalies.These methods neither utilize high-level features and latent spatial ones,nor the diversities of both normal and abnormal patterns.To overcome these drawbacks,we propose three unsupervised anomaly detection models by using deep learning-based methods.The contents and research results are as follows.First,we review the current research status of anomaly detection and point out the problems to be solved at the present stage.At the same time,we introduce basic theories of anomaly detection based on deep learning and two commonly used anomaly detection frameworks,i.e.,autoencoders and generative adversarial networks.Next,aiming at the defect that existing methods ignore the features in the latent space,we design a latent-feature autoencoder via adversarial training(LAA),and propose a weighted feature consistency loss and feature discrimination loss.LAA utilizes the features in the latent space of the autoencoder and the discriminator and achieves 85.1% and 80.6% on the CIFAR-10 and CIFAR-100 datasets in terms of the area under the receiver operating characteristic(AUC),respectively.The detection speed on UCSD Ped2 dataset can reach 60 frames per second(FPS),which fully reflects the important utilization value of latent spatial features.Then,to combine latent spatial features with temporal ones,we propound a predictionbased method named an autoencoder with a memory module(AMM)and a memory triplet loss.The input of AMM is a sequence of previous video frames,while the output of AMM is a future frame.We unite a prediction loss and multi-scale structure similarity measure to detect anomalies with multi-views and multi-scales.The AUC and detection speed of AMM on UCSD Ped2 dataset reach 97.2% and 75 FPS,which can be used in the intelligent video surveillances.Finally,based on the above models,we study the event completion-based video anomaly detection method and propose a multimodal event completion autoencoders(MECA).MECA employs a pre-trained object detector and gradient images to produce incomplete video events and then combines an appearance autoencoder and a motion autoencoder to complete incomplete video events in appearance modality and motion modality,respectively.MECA simultaneously utilizes latent spatial features,temporal features,motion features,and high-level semantic information.AUC on UCSD Ped2 and CUHK Avenue datasets reached 97.8% and90.8%,respectively.Comprehensive experimental results on four benchmark datasets demonstrate that MECA is effective in video anomaly detection and outperforms several stateof-the-art methods,such as MNAD,Mem AE,and VEC.