Research On Crowd Collectiveness Recognition Based On Deep Learning

Group motion pattern recognition is very important to crowd management in public places.It can provide decision support for crowd emergency identification,crowd evacuation,safety management in public places and many other application fields.The collectiveness of crowd movement not only depends on the individual,but also is affected by the local movement state of the crowd.When the local collectiveness is inconsistent with the global collectiveness,it will lead to the wrong identification of the crowd collectiveness,and then affect the relevant decision making.In addition,due to the complex structure and large number of parameters of the crowd collectiveness recognition model,the application of crowd collective safety monitoring through the model will lead to resource occupation and slow response,which will affect the overall performance of the system.In view of the above problems,this paper carries out relevant research on the crowd collectiveness.The main contributions of the paper are as follows:(1)This paper proposed a crowd collectiveness recognition Convolutional Neural Network,which combines the local features and the global features.Firstly,the collective measure images are constructed based on the optical flow vectors in this algorithm.Secondly,channel attention is added after the first layer of network convolution to obtain the global information on crowd motion,and the dilated convolution is used to obtain the local crowd motion information.This method solves the problem of the uniformity between the global and local collectiveness of the crowd in public places.(2)This paper improves Res Ne Xt network as distillation of knowledge teacher network,replace convolution kernels for dilated convolution kernels and adding channel module to obtain global and local feature information.Then simplify the crowd collectiveness convolution network as a student network.The knowledge distillation network is proposed for the crowd collectiveness recognition,which improves the overall performance of the model.Relevant experiments are carried out on the WWW crowd dataset of public places.The weighted average recall,weighted average accuracy and weighted average precision of the crowd collectiveness convolutional network recognition method combining local and global features are92.4%,94.9% and 92.6%.The collectiveness recognition knowledge distillation networks increased by 0.7%,0.9% and 2.6% on this basis.The experimental results show that two proposed methods achieve better performance of crowd collectiveness recognition in crowd scenes.