Algorithm Of Crowd Density Estimation Based On Gray Level Co-occurrence Matrix

For the impact of global urbanization，dense crowds moving and rendezvousing wouldeasily cause serious trampling accidents. These unfortunate accidents happen frequently in theworld. At the same time, the increase of crowd density will make the city’s public transportface short-term peak. If the congested crowd can not be evacuated in time, it will pose agreater threat to urban public security. Thus, automatic analysis of the crowd information andestimating of the crowd density have become a research focus in the intelligent videosurveillance.The focus of this thesis is how to estimate the crowd density of video surveillance.Firstly, extract the crowd foreground image of video image in monitoring scene, and thenextract the crowd density characteristics of the foreground image. Finally, the crowd densitycharacteristics are sent to classifier, thereby obtaining the crowd density.In the aspect of crowd foreground extraction, the weighted average method is adopted togray the crowd image, then the median filtering method is used to eliminate noise, and finally,adaptive change detection mask method is applied to construct the background image, therebygetting crowd foreground image by background subtraction operation.As to the crowd density characteristics extraction, the development and basic principlesof crowd density monitoring system both in China and abroad is introduced. Thetexture-based analysis method is very effective for the high-density scene. Therefore, thetexture analysis method based on Gray Level Co-occurrence Matrix is used to extract thecrowd density characteristics. The best parameters of GLCM are determined by experiments,and the four important characteristics of the energy, entropy, contrast and homogeneity areselected as the texture features of crowd image.Regarding pattern recognition classification, the support vector machine (SVM) isadopted as classifier to estimate the crowd density. According to the classification rules, thesupport vector machine classifier model is established by use of training samples. By experiments, The coarse and fine tuning method is used to test and study the selection ofthe best combination of the radial basis function’s kernel parameter and penalty parameter C.Finally, in order to verify the effectiveness of the algorithm, experiments are carried outon two different crowd videos. The accuracy rate of test samples reaches more than95%. Theexperimental results show that the method is simple and effective with convenience forapplication in the actual scene.