Crowd Motion Pattern Learning And Anomaly Detection Based On Surveillance Data

In recent years,as people's awareness of safety prevention and control has increased,more and more surveillance equipment have been installed in various public spaces such as streets,shopping malls,gymnasiums,and office buildings.These surveillance data contain massive amounts of information that are urgently needed selected and screened,coupled with the continuous development of computer technology,the mining and analysis of crowd trajectory data has gradually become a popular research direction.It involves many aspects such as security protection,public space design,and event planning,which are worthy of in-depth research.This paper focuses on the crowd trajectories from the surveillance data and analyzes the crowd behavior,mainly from the three aspects of motion pattern learning,abnormal trajectory detection and trajectory prediction.In terms of motion pattern learning,first,the Structure Preserving Compound Algorithm(SPC)is proposed to preprocess the trajectories,the trajectories are encoded,and then the dimension is reduced to facilitate the subsequent clustering algorithm computing.And then the AEM-LDA algorithm(Annealing-based Expectation Maximization for Latent Dirichlet Allocation)is proposed to cluster the trajectories.Aiming at the problem that the EM algorithm may fall into local optimization during the LDA model parameter inference process,simulated annealing is used to improve it,which improves the accuracy of model training and obtains trajectory clusters.Finally,the trajectory clusters are analyzed to obtain two types of regions of interest,namely Optimum Path and Critical Regions,which are used to describe crowd motion patterns.In terms of crowd trajectory prediction,the existing neural network prediction algorithms directly predict a definite result,which cannot reflect the uncertainty in the prediction.This paper improves the conv LSTM neural network.By adding the Mixed Density Network(MDN)layer to the network,the linear combination of Gaussian kernel function is used to simulate the probability distribution of the output,rather than using a uniquely determined position,the accuracy of the trajectory prediction effectively improves.In terms of anomaly trajectory detection,to solve the problem of single detection standard using anomaly detection algorithm alone and low detection accuracy,this paper proposes a new anomaly detection framework.First,Isolation based Anomaly Detection(IAD)is proposed to classify the trajectories and perform anomaly detection within the classes.After obtaining the preliminary detection results,and then combined with the Edit Distance on Real Penalty(ERP)to perform trajectory similarity measurement to further identify whether the trajectory is abnormal,which can effectively improve the accuracy of anomaly detection and reduce the false alarm rate.