Research Of Crowd Size Estimation Method Based On Image Processing

Recently, with development of modern life, kinds of social activities become morefrequent. It’s naturally that a range of people participate in these activities which need highdemanding of effective management and protection of crowd safety. A number of stampedesaccrued in these years make people understand of the importance of monitoring for the crowd.The traditional manual surveillance systems has many deficiencies, on one hand it cost hugeof human resources, on the other hand the observers could not always stay focus. Additionally,psychologist points that there are some limitations of humans ability for monitoringsimultaneous signals[1]. So the intelligent video monitor with automatically crowd analysisand identification has a high practical value. A clear relationship between level of services andpedestrian flow is provided by Polus et al[2]. It’s naturally that crowd with high densityrequires high level of services. Crowd density is an important measurement of crowd modelswhich make automatically crowd density analysis become a hot topic.There are a lot of research papers on this topic, and many constructive algorithms areproposed. Existing literatures targeted crowd size estimation in a certain surveillance areamainly divide into two types: The first category is direct count based detecting and separatingby individuals. The second category is indirect count based crowd features and machinelearning. Direct count often use shape indexing, face detection, skin color, motion, features,and only heads or the shape formed by heads and shoulders to detect and segmentindividuals. The second type algorithms regard the crowd as whole features, and then build arelationship between the feature vector and crowd density. The indirect count method has abetter robustness in the complex background as not rigidly stick details.In this paper, we use indirect method to estimate crowd density. The indirect methodmake up by feature extraction and classifier design. The nature of feature extraction isreducing dimension which let us use fewer dimension to describe the crowd characteristics.Duo to the location of camera, capture frames influenced by perspective distortion make thepeople show different scales. Moreover, when it comes to high density, overlap among crowdmembers can be solved by using texture analysis. So we take a multi-resolution textureanalysis algorithm. Actually, we adopt two multi-resolution analysis tools: Dual-TreeComplex Wavelet Transform (DT-CWT) and Gabor filters. Crowd density analysis is a kindof multi-class pattern recognition problems. Crowd density estimation based on textureanalysis is a kind of crowd texture classification. Feature vector extracted from crowd imageis sent to classifier to determine the final classification. As there are not very much experimental samples, we use support vector machine as our final classification tools.Compared with the traditional real wavelet transform, DT-CWT has better directionaland approximate translation invariance and has been widely used in texture segmentation andclassification. Here we use DT-CWT and18high-frequency sub-bands and2LL sub-bandswill obtain after three-layer decomposition. Texture features are taken from these sub-bands.We use energy, anisotropy of rotation invariant features and automatic correction factor tobuild our texture vector respectively. We use PETS2009dataset to build ourselves’benchmarkdata. Experimental results show that DT-CWT has fine performance. Another tool is Gaborfilter which is widely used in texture classification and texture segmentation. The key point isdesigning appropriate Gabor filter bank. These filters are used as convolution with crowdimages, a statistical analysis is used to extract features for filtered images. In this paper, weuse three different sets of Gabor filter channels: even symmetric multichannel Gabor filters,odd symmetric multichannel Gabor filters, the visual cortical channels formed by a pair ofeven-and odd-symmetric Gabor filters. Experiment indicate that the odd Gabor filter havebetter performance by considering the amount of computation.