Research On The Algorithm Of Abnormal Group Activities Detection Based On Videos

Group abnormal behavior detection as a key technology in the field of video surveillance has important research value.Timely and accurately discovery abnormal behaviors in the population,such as abnormal gathering,fighting,etc.,and then taking measures to maintaining public safety and safeguarding the lives and property of the people is very important.With the rapid development of machine learning,especially deep learning,image-based individual abnormal behavior detection technology is more and more mature,but video-based group abnormal behavior detection technology still has many difficulties to be solved,such as poor real-time performance and low detection accuracy.This paper first introduces the relevant knowledge of group anomaly detection,including the related theories of traditional machine learning and deep learning,and then the traditional machine learning and deep learning methods are used to detect the abnormal behavior of the video in the UMN dataset,UCSD dataset,and self-built dataset,and the experimental results were compared with other algorithms.The work of this dissertation is:1.The basic theory of group abnormal behavior detection is introduced from the aspects of traditional machine learning and deep learning.In the aspect of traditional machine learning,the process of group abnormal behavior detection is introduced,the basic principles of sparse optical flow method and dense optical flow method are introduced too,and then the two-class model support vector machine is introduced.In the aspect of deep learning,the concepts of deep learning,the structure of neural network and its training process are introduced.The autoencoder model and convolutional neural network model in deep learning model are introduced in detail.2.A group abnormal behavior detection algorithm combining complex wavelet domain denoising and particle swarm optimization twin support vector machine(PSO-TSVM)is proposed.Firstly,the speed,acceleration,directional characteristics and population density characteristics of group behavior in video are extracted by Horn-Schunck optical flow method,and then the non-subsampled dual-tree complex wavelet packet transform(NS-DTCWPT)is used to perform multi-resolution decomposition on the extracted features,and the bivariate model is used to remove the noise in the extracted features.NS-DTCWPT achieves finer frequency band division in the whole frequency band and has translation invariance.The two-variable denoising model considers the correlation between different wavelet coefficients after multi-resolution decomposition of the signal,which improves the denoising effect.Twin Support Vector Machine(TSVM)has a good processing ability for unbalanced data,the twin support vector machine model optimized by particle swarm optimization has better generalization performance.Experimental results on UMN video datasets and self-built datasets shows that the proposed algorithm has higher detection accuracy than the social force model and particle entropy model.3.A group abnormal behavior detection algorithm based on multi-column spatial-temporal autoencoder is proposed.The algorithm uses a simple and effective multi-column spatial-temporal autoencoder network architecture to input video images as input,and automatically capture the spatial structure in the data with the help of an autoencoder,these spatial structures are combined to form a video representation,then the representation is fed into the stack of multi-column spatial-temporal autoencoder to learn the regular time pattern.The proposed multi-column spatial-temporal autoencoder allows the input image to have an arbitrary size or resolution,by using filters of different sizes to sense the field of view,the features learned from each column of the encoder can be adapted to changes in the size of the anomalous events in the video due to perspective or image resolution.At the same time,the three-layer convolution long short term memory LSTM model is introduced.LSTM can solve the problem that RNN can not effectively use the long-span historical information,and avoid the problem of gradient disappearing during model training.