Research Of The Dense Crowd Counting Algorithm Based On Convolutional Neural Network

In recent years,monitoring of crowds has become particularly important because of the rapid population growth and the frequent occurrence of riots caused by crowds.The method of computer vision used to estimate the density and behavior of the crowded population has become a popular research direction,and has a wide range of practical application value.In this thesis,we study the two aspects:population density estimation and people number estimation.The density estimation is to classify the population areas to high intensity,medium intensity,and low density levels.Most of the conventional research methods are based on the following framework:1)extract the features of the population area,such as texture features or wavelet features.2)Estimate the number of people by means of target detection or regression algorithms.However,in dense scenes,the distribution of individuals in the population is diversified,coupled with the targets occlusion in the complex environment and light changes,where the high-performance feature representation and people number estimation are difficult.In recent years,the deep learning has been successful in the field of computer vision.The related research shows that the features learned by deep neural network are more general and representative than traditional visual features.Motivated by the previous research work,this thesis proposes a novel crowd people number estimation and population density classification algorithm based on deep convolution network.Through the use of two supervised signal in learning,the proposed method improves the robustness of network learning,inhibit the over-fitting which may occurs in learning process.The main contributions of this thesis are summarized as follows:1)We propose a convolution neural network to learn the features of the population in the image area.Through learning by the network,the population in image regions can be divided into different density levels,to fulfill the task of population high-density early warning.2)We proposed a two-way monitoring signal in model learning.One signal is for the area population density classification,and the other is for people number counting in the area.The two-way learning method increases the robustness of the network learning and improves the accuracy of the estimated people number.3)We have created the data set AHU-CROWD,together with the current commonly used public data sets,including UCF-CROWD,UCSD,to perform experimental verification.The results show that compared with the previous methods,our proposed method has a significant performance improvement in crowd population estimation.