The Large Scale Crowd Analysis Based On High Definition Video

Along with the rapid development of economy and the society , large-scale crowd activ-ities are ever increasingly frequent. Particularly, the scope and number of public events, forexample competitive sports, recreational activities and various exhibitions, expanding causesthat crowd security monitoring becomes the closely-attention hotspot of the relevant depart-ments. The growth of video monitoring and its correlation techniques lays the foundationfor crowd management and supervisory from the point view of video analysis. Crowd ?owand density are two important indexes of crowd monitoring, and also the critical basis ofcrowd management. Polus[1]proposes that crowd density is related to crowd safety and levelof service that crowd acquires. Such theory plays an important role to the research on therelationship between crowd density and its safety.Traditional video monitoring is depending on human operation which is a waste of timeand energy, and in?icts situations like false-alarm or false dismissal. Recently, as the devel-opment of computer visions, artificial intelligence and correlation techniques, algorithms ofcrowd density estimation and ?ow statistics based on intelligent video analysis are emerg-ing. However, existing algorithms of density analysis and ?ow statistics are founded onseparation technology, which will limit the application of system when the complexity in-creases. The thesis focuses on density analysis and people counting for large-scale crowd,summarizes and compares the mainstream algorithms of crowd density analysis and crowdstatistics. Based on the previous working, the author further works on the following aspects.Firstly, the thesis carries out and improves the crowd feature pre-processing technologyfor large-scale crowd analysis. During the process of people feature extraction, both videoshaking and the existence of shadow do great negative effect to system precision. The au-thor improves the algorithm of foreground exaction based on background subtraction usingwidow method. Then realizing shadow elimination of foreground by color-space transform.In the procedure of crowd density analysis and crowd statistics, the perspective effect broughtby different distance between people and video tremendously in?uents the precision of thesystem. In this case, the author effectively solves the perspective effect caused by crowd feature distortion using perspective correction algorithm according to linear interpolationweight.Furthermore, the thesis improves and proposes some algorithms of crowd feature ex-traction. On the one hand, the author solves perspective effect in the process of peoplefeature extraction by introducing perspective correction index in index statistic feature andGrey Level dependence matrix. On the other hand, the author proposes the algorithm ofpeople feature extraction based on Maximally Stable Extreme Regions for the multi-viewfeature problem, which is caused by video rotation in monitoring, installation and positiondifferences of camera. The Maximally Stable Extreme Region is invariant to affine trans-formation, thus has better adaption in the situation of multi-views. At the same time, theauthor proposes the crowd feature extraction algorithm based on Gabor filtering dictionaryfor the precision requirement of crowd statistic system. Gabor filtering dictionary can pro-vide multi-direction, multi-scale information of crowd feature which is more adaptive forlarge-scale crowd estimation.Lastly, the thesis designs the system of crowd density estimation and people countingbased on the key technologies introduced previously. In the first place, the author compre-hensively utilizes the pixel statistic feature and improved Grey Level Dependence Matrixto analyze the crowd density, since they perform separately well in low density region andhigh density region, respectively. Therefore, the proposed system enjoys very high precisionin full density region. When the view changing caused by video rotation exists, the authordesigns a system of crowd density analysis based on improved Maximally Stable ExtremeRegion, which improves the precision of the system for its affine invariance. In the aspectof crowd density analysis, the thesis designs a people counting algorithm depended on thecombination of Gabor Filtering dictionary and Regression. Then the author establishes thefunction of crowd feature and in-sight people using vector machine regression technologyto count the people. Comparing with traditional people counting algorithms with extraction,the paper focuses on the integral rather than individual, which is more suitable in large-scalepeople counting.