The Algorithm Of Crowd Density Estimation And The Application Of Crowd Behavior Analysis Based On Deep Learning

With the rapid and increasing mobility of society,many crowd violence event and crowd tread event had been increased day by day.Crowd Density Estimation(CDE)and Crowd Behavior Analysis(CBA)algorithms have become a hot points in computer vision(CV).From now on,the CDE has a lot of problems need to be solved,such as keep out,scale variation,light intensity variation and low resolution and so on,so the methods of analyzing individual for CDE and CBA can't work in the hard environment.Besides,the traditional shallow learning models have large false drop rate in CDE and CBA,but the deep learning models have been a hot research point in recent years.So,in this article,we have analyzed the existing algorithms for CDE and CBA's defects,extended the traditional features and constructed a deep Convolution Neural Network(CNN)for training and testing,and this indeed improves the accuracy and robustness of CDE and CBA.Our works in the article as following:1.We have reviewed the theoretical basis for CDE and CBA.We review the color feature,texture feature and contour feature in CDE,and the spatio-temporal interest points feature and optical flow feature in CBA,especially for the two frequently used algorithms in CDE and CBA,BP networks and Support Vector Machine.In the machine learning and patterns recognition algorithm for CDE and CBA.Then,we extend the machine learning model to deep models,three common deep neural network have been illustrated,they are CNN,RNN and DBN.In the following we summarize the researches in the CDE and CBA based on deep learning models,laying the base for the further research.2.The algorithm of CDE based on deep CNN model.In order to improve the accuracy and robustness of CDE algorithm and monitor the crowd density in real time,we extracted the features by the Complete Local Binary Patterns(CLBP)algorithm for density texture description that belong to improved LBP features.The features include the difference magnitude features,central pixel gray level features and LBP sign features.After that we used a deep CNN for CDE detection,convolution operations,pooling operations,and Back Propagation(BP)algorithm are used for training the neural network.Experimental results show that,CLBP and CNN can adapt to complicated episode,and the deep model gives the essential features,this will result in the recognition of CDE has been improved,and the robustness is also better.3.The CBA analysis is based on the deep CNN,Firstly we compare the crowd group object extraction algorithms,and adopt the PBAS method to extract group crowd in images.Then we extract group motion features on group crowd.These motion features include collectivity,stability and conflict features,they can display the essential features for crowd's abnormal behaviors,and construct high recognition in the CBA.Then a deep CNN is applied in CBA.Experimental results show that the motion features can extend from independent episode,for any episodes they can give a robust result,and the CNN deep model has increased the accuracy of detecting crowd's abnormal behaviors.