Crowd Density Estimation And Crowd Abnormal Human Behavior Recognition In Video Surveillance

Due to the development of image processing,pattern recognition and artificial intelligence,intelligent video surveillance has gradually shifted from the theory to the stage of widespread application.Therefore,intelligent video surveillance technology as the core technology to improve traditional monitoring system has been pay more and more attention.And the crowd analysis including the crowd density estimation and crowd behavior recognition which as one of the great challenges in intelligent video surveillance has become a major research goal of researchers.Based on the detailed review and analysis of the crowd and the relevant information,we implemented research for some relative specific problems,such as label distribution learning,multi-label learning,crowd density estimation,the crowd behavior classification,the main research content of this paper is as follows:1.This paper introduces basic theory knowledge about the multi-label learning and label distribution learning.And elaborates the two main multi-label learning algorithm methods that problem reforming and algorithmic adaption.In order to obtain an accurate classification model,current research on multi-label classification also assumes that large amounts of labeled training data are existing.For this shortcoming,transductive multi-label learning is proposed.And comparing to other methods,the advantages of transductive multi-label learning are shown.Label distribution considers that each crowd image corresponds to the label distribution,and a label distribution is composed by the degrees that every label describes the image.Label distribution learning considers all the possible label in the learning process,which enhances the generation capacity of the classifier.2.This paper proposes a crowd density estimation algorithm based on the label distribution learning.In this paper,for each frame of the crowd image is assigned a probability distribution to describe the degree of the image represent the number of labels.Firstly,we deals with the crowd image pixels are by a method of perspectivenormalisation,segments the crowd in the video through the algorithm of crowd segmentation,and extracts crowd image low level features include segmentation feature,texture feature,edge feature;then use conditional probability neural network learning algorithm to train a number of label distribution model;Finally,the number of people in the image can be obtained when input a test image to the model.In addition to evaluate the performance of the algorithm by the number of errors,we also compared the predicted label distribution of test samples and the true label distribution.Experimental results show that our approach can perform remarkably better than the compared to other crowd density estimation algorithms on UCSD data sets.3.This papers put forward two kinds of crowd abnormal behavior recognition algorithm.We consider the abnormal crowd behavior recognition as a multi-label classification problem,every behavior sequence can be assigned one or more labels.And think effective classification model is obtained by labeled or unlabeled behavior sequence information,then complete the identification of the abnormal crowd behavior.The algorithm of Abnormal behavior recognition algorithm based on preference of distributed learning is that through learning inconsistent rankings to obtain preference distribution which can be compatible with multiple inconsistent rankings,then we use the BFGS(Broyden Fletcher goldfard Shann)label distribution learning algorithm to learn the latent preference distribution model,finally completes the behavior classification,this method solve problem that label ranking is inconsistent.Abnormal behavior recognition algorithm based on transductive multi-label will use a large number of test sets of data information in the multi-label learning algorithm,which make the algorithm obtains better classification results behavior on test data set.Experimental results show that the two crowd abnormal behavior recognition methods perform remarkably better than the compared state-of-the-art algorithms.