Research On Activity Analysis And Semantic Retrieval In Distributed Intelligent Visual Surveillance System

With the development of network multimedia technology and the increasingneeds in different surveillance scenarios such as public,commercial and education,visual surveillance has become a major application of computer vision and is of greatsocietal importance.The current generation of surveillance system aims to provide adigital,distributed and intelligent solution.Currently,intelligent visual surveillancehas become a research focus of computer vision and involves many research areaincluding machine learning,pattern recognition,artificial intelligence and data mining.Different from the traditional video surveillance system,intelligent visual surveillancesystem can accomplish the entire surveillance task as automatically as possible.Theprocess consists of automatically acquiring information by cameras,providinginterpretations of scenes,understanding the actions of objects and analyzing the videosequence without or with less help of human beings.This paper focuses on the problem of behaviors understanding and semanticretrieval in visual surveillance system and proposes the architecture of the distributedintelligent visual surveillance and retrieval system.Firstly,the behavior understanding and classification algorithm are studied and asub-trajectory description based trajectory feature extraction algorithm is proposed.Sub-trajectories are sampled to reveal the local activities of trajectories.Thedistribution patterns of sub-trajectories are then constructed by learning.Therefore,the distribution patterns of sub-trajectories can be used to represent and extract thefeatures of trajectories.Based on the trajectory feature extraction algorithm proposed,an anomaly detection and prediction algorithm is implemented using SOM,on theother hand a trajectory classification algorithm is implemented using frequent patternsmining.Secondly,the semantic retrieval algorithm in intelligent visual surveillancesystem is studied and a distance preserving transformation based similaritymeasurement algorithm is proposed.The similarity of trajectories is measured bymulti-dimensional Hausdorff distance.Distance preserving transformation is used tooptimize the similarity measure which reduce the computational cost and provide anew trajectory descriptor.Based on the similarity measurement algorithm proposed,atrajectory clustering algorithm is implemented by spectral clustering,on the otherhand a SVM based interactive semantic retrieval algorithm is implemented using relevance feedback.Thirdly,the semantic retrieval algorithm in distributed intelligent visualsurveillance system is studied and a metric space based distributed interactiveretrieval algorithm is proposed.The feature vectors of objects are distributed to astructural peer to peer network by M-Chord.The metric spaces generated by SVMand M-Chord are then combined.Therefore,the retrieval performance can beimproved by the pivots in the metric space.Furthermore,a load balance scheme usingthe concept of virtual servers is used to achieve load balance in the heterogeneous anddynamic surveillance network.Then,the subspace clustering algorithm in distributed intelligent visualsurveillance system is studied and a distributed voting based subspace clusteringalgorithm is proposed.The problem of distributed subspace clustering is converted tothe distributed voting problem.It used the topology of the underlying overlay networkto collect the voting information hierarchically.Therefore,all the clustering results ofnodes can be collected by distributed voting.According to the speciality of subspaceclustering,the dataset reducing,pruning and region merging schemes are introducedto further optimize the communication between nodes.Finally,the architecture and framework of large scale metropolitan intelligentvisual surveillance system is studied and the prototype of the distributed intelligentvisual surveillance and retrieval system is constructed.The topology of the distributedvisual surveillance and retrieval system is studied.The prototype of embeddedsurveillance terminal is introduced.The data flows and transaction flows in the systemare analyzed.