Research And Design Of People Counting Algorithm Based On The Combination Of Cloud And Terminal

With the improvement of urban living standards,citizens' attention to public security continues to increase.However,in recent years,Safety accidents caused by large scale crowd gathering activities are not uncommon.Therefore,it is very meaningful to use the existing public monitoring systems to estimate the number and intensity of crowds and the potential risk of safety accidents in a crowded scene,and to provide information for artificial crowd management.In recent years,the field of people counting based on c crowd counting has been developing continuously,and the existing people counting methods have reached a high accuracy rate.However,these algorithms need the support of hardware devices with powerful computing resources,and the reasoning time of algorithms is long,which cannot meet the real-time people counting requirements.Based on the idea of the combination of cloud and terminal,this thesis designs two methods to count the number of people in crowded scenes.At the same time,based on the two methods of people counting designed in this thesis,an intelligent people counting system based on the combination of cloud and terminal is developed,which can provide real-time people counting service for crowd monitoring and management in crowded scenes.This thesis main content is as follows:(1)To work out the issue of bad accuracy of people counting method due to the big head size change in crowded scenes,a people counting method called MSANET was proposed.This method aims at the characteristics of the number of people in the scene of high crowd density,large change of head size and serious occlusion,a atrous spatial pyramid module for crowd density estimation was proposed,and the dilation rate used in the dilated convolution operation of the module was elaborate formulated to get the multi-scale information in the crowd feature map,so as to improve the processing ability of human head scale change.At the same time,attention mechanism is introduced to optimize the background part of the "middle" feature map of the model by the application of attention mask.The experiments indicate that MSANET,the people counting method based on crowd counting,can better deal with the change of head scale and improve the accuracy of the people counting algorithm.(2)The existing people counting methods require high computational resources,take a long time to calculate,and can't meet the real-time people counting needs.In this thesis,a lightweight people counting method based on crowd counting,DDSNET,is proposed.In this method,the encoder of the model is constructed using the dilated depth-wise separable convolution module proposed in this thesis.Compared with the standard convolution calculation,the dilated depth-wise separable convolution module significantly decreases parameters.The experiments indicate that compared with other people counting methods,DDSNET can obtain better prediction accuracy with less parameters,lower computational complexity and faster speed.(3)MSANET method and DDSNET method proposed in this thesis are used as the core algorithm of people counting.Based on the mode of collaborative processing between cloud and terminal,a people counting system based on the combination of cloud and terminal is designed and developed,which provides important reference information for crowd supervision in high-density scenarios.