Crowd Density Estimation Based On Convolutional Neural Network

With the development of economy and fast growth of population, the disturbances caused by the crowd have increased. Based on the fact of that, crowd management has gained more and more attentions. But monitoring crowd only by manual work is likely to be affected by fatigue and personal will. At the same time, computer vision technology has been widely used in vehicle license recognition, face recognition and finger print recognition, which also enhances the research of automation for crowd management and control.Crowd density estimation and crowd flow surveillance are two important directions for crowd management. Crowd density estimation is often used for precaution, in case of the crowd extends the safety margin. The crowd flow surveillance focuses on people counting and is often used in some applications which need to understand the exact number of people. Crowd estimation contains two steps: feature extraction and crowd density classification. The developed methods exist some problems: To promote the effect of crowd density classification, researchers developed many complicate feature extraction methods which, as a result, lowers down the detection speed. Besides, the classification methods such as SVM are shallow learning methods which contain limitations to some extent.Recently, as the developing of deep learning, people put high value on its multiple layer structure. As a representative model of deep learning, deep convolution neural network extracts high level features based on its multi-layer structure through feature learning. In this thesis, we introduce convolution neural network into crowd density estimation and make some innovations on the network structure and the main works are as follows:1) We analysis 3 feature extraction methods and 2 classification methods and we introduce convolution neural network into crowd density estimation.2) We propose a fast method of crowd density estimation based on convolution neural network. We optimize the structure of convolution neural network to speed up the detection and further develop a cascade optimized network to improve the accuracy. Finally, we apply the optimized model into crowd density estimation and conduct many contrast experiments which prove the efficiency of our method.3) We develop a crowd density surveillance system based on the improved method. Considering the characteristic of crowd image, we add geometric calibration into our system to improve the effect.