Research On The Methods Of Crowd Density Estimation And Motion State Analysis Based On Video Sequence

Crowd counting in public places, such as squares, stations and other places, is an important subject in the field of public security, it is of important significance and widely application value in aspect of intelligent information processing. As a key technology of intelligent monitoring, intelligent crowd monitoring system has become one of hot topics and it has great practical significance in crowd control. How to achieve good performance of crowd analysis is still a problem being further researched. This dissertation employed the wavelet decomposition in the crowd analysis and proposed an evaluation algorithm of crowd number estimation and motion direction distribution. In frequency domain, crowd number and turbulence of the crowd motion were used for the detection crowd abnormity.First of all, according to the good anti-noise performance of video image processing in frequency domain, a crowd number estimation algorithm is proposed based on frequency-domain energy. The video image is decomposed into two levels by two levels wavelet decomposition with the algorithm, and then high-frequency detail sub-band image is used to extract the energy of the crowd population. The relationship of crowd number and energy is used to gain the number of people by the method of linear regression.Then, the block matching algorithm is applied to estimate the direction of population motion from the approximate image obtained by wavelet decomposition. The motion vector of the crowd feature point is figured out to the motion direction histogram. In this motion direction histogram, the motion direction of the crowd can be described clearly, and then the information entropy achieved from the motion direction histogram is employed to express the confusion level of the crowd.Finally, the crowd state was judged by combining the crowd number and information entropy of motion direction distribution. The current state of crowd could be more practically reflected by the combination of crowd population and crowd motion state. For the gathering and scattering of the crowd, two aspects of crowd number and motion trend were analyzed and researched in this dissertation.The experimental results show that compared with the algorithm which uses the crowd density or the motion feature only, the proposed detection method of crowd abnormal in the thesis has characteristics of simple mathematical models and high accuracy. In addition, feature extraction in the frequency domain has a reasonably robust to noise.