The Study On Crowd Density Estimation

The rapid development of computer vision and artificial intelligence lays a solidfoundation for intelligent analysis and monitoring which are based on the viewpoint ofvideo. Then they provide technical supports to public security alarming field and crowdchannel field. Crowd density estimation is the key to these fields.In this article, we make a comprehensive and systematical description of the crowddensity estimation algorithms and launch the research from feature pretreatment, featureselection, algorithms process improvement, enrich samples and experimental contrastlink. In each link we have obtained the corresponding results.In feature pretreatment link，we have improved the shadow suppression algorithmwhich is based on the color feature invariant. It enhances the effect of shadowsuppression.In feature selection link, we study and summarize the current mainstream algorithmsof crowd density estimation by reading vast amount of literatures. We also make acomparative analysis of the advantages, disadvantages and the scope of three kinds ofalgorithms which are based on pixel-based analysis, texture-level analysis andobject-level analysis. In this article we propose an algorithm which combines thecharacteristics of pixels and texture analysis. Because it is relatively straightforward toutilize prospect area feature to do the image classification, however, it is more meticulousto utilize fractal dimension. So we extract the image characteristics of fractal dimension,prospect area and edge pixels on prospects as the feature vector. The algorithmovercomes the defect of highly depends on viewpoints which precisely is thedisadvantages of algorithms based on the overall texture analysis. This can enhance thegeneralization performance of algorithms. Because of the multi-scale analysischaracteristics of fractal dimension, which means there is no need to implementperspective correction. This can reduce the complexity of the algorithm in using process.The experiment research showed that the combination of fractal dimension and prospectarea can play a complementary role in the scheme, which increases algorithm analysisability. Meanwhile, the feature dimension of the algorithm is simple and characteristicanalysis is relatively easy. In a nutshell the algorithm can be intuitively, real-timely response the number and density of people in practical.We use the correlation of video on the time-series and add the time series correctionon the part of the algorithm flow of traditional crowd density estimation. Time seriescorrection can reduce the false alarm rate and miss rate effectively，also can analyze theabnormal conditions. Thus, it has important application values in the crowd densityestimation and early warning alarm.This article not only use two public libraries which are PETS2010and UCSD, butalso use campus monitoring data to construct the rich variety of sample photo librariesand video libraries. We get three different climate conditions’ videos which are the rainy,the snowy and the sunny day in the view of monitor of activity center. We aims to findthe commonness and differences of the characteristics at different scenarios or the samescene under different natural state. Comprehensive experiments for perspective effectcorrection, classification ability and generalization performance are used to ensure theactual reliability of the experimental results.In order to verify the validity of the algorithm in this article. We compare with thecurrent typical algorithms such as pixel features, Gabor texture, the Wavelet packet,Minkowski dimension and Gray symbiotic matrix.