Research On Abnormal Event Detection In Video Surveillance

With the increasing prominence of public safety issues,timely detection of abnormal events in public places will help the relevant security departments to take actions and provide rescue quickly,and thereby reduce personal injury and property losses of residents.In recent years,with the rapid development of intelligent monitoring and security technology,video-based anomaly detection has become one of the hot research topics in image processing,machine vision,machine learning and some other fields.The main task of abnormal event detection is to automatically filter the abnormal events in videos,and then feedback and address these abnormal events.But in practice,it is not feasible to enumerate all abnormal situations that may occur in a given surveillance video sequence,which makes the research more challenging.In addition,abnormal events in complex video provide a lot of valuable information for intelligent surveillance,automatic alarm system and other forward-looking applications,which makes this problem more realistic.At present,all the frontier methods to solve the problem of abnormal event detection can be roughly divided into two categories.One is based on the algorithm using shallow hand-crafted feature extraction,and the other is based on the algorithm using deeply understanding feature extraction.Most of these two kinds of methods adopt the way of piece-wise learning,which will inevitably lead to local optimal solution problems.In this thesis,a novel solution for video anomaly detection is proposed,which integrates oneclass support vector machines into Convolutional Neural Networks,that is,an end-to-end deep one-class anomaly detection model.Furthermore,a robust loss function derived from one-class SVM is introduced to optimize the parameters of the model.Compared with the piece-wise models,our detection model not only simplifies the complexity of the process,but also obtains the global optimal solution of the whole process.The experimental results show that our deep one-class detection model has good performance and is very effective for abnormal event detection in surveillance video.In addition,most of the existing research algorithms cannot deal with small-scale data and large-scale high-dimensional data at the same time.Some methods of processing small-scale data will show super-high computational complexity when applied to high-dimensional data,while some methods of processing high-dimensional data will take up a lot of storage space.In order to solve these problems,a sparse representation framework of hidden spatial dictionary based on Variational Auto-Encoder is proposed.For large datasets,the method can reduce dimensionality,get hidden information and extract more meaningful features than shallow hand-crafted features.Meanwhile,for the storage of normal event information,the space cost is greatly reduced.In order to verify the generality and performance of the algorithm,the experiment is conducted on different types of anomaly detection tasks.In addition to UCSD pedestrian dataset used in the above algorithm,KDD-CUP'99 dataset for network intrusion detection and Mnist dataset for image anomaly detection are also employed.The experimental results show that the algorithm has great experimental results in various anomaly detection tasks.