Lightweight And Feature Sharing Convolutional Neural Network For Dense Crowd Image Analysis

With the development of economy and the advancement of society,the scale of world population is expanding,the density of population is increasing,and the places and forms of human social activities become richer,at the same time,the safety of crowd has become the focus of attention from all sectors of society.Some public places such as shopping malls,gymnasiums,concert halls,commercial streets and so on are often distributed with different scale of crowd.With the increase of the number of crowd,the probability of safety accidents increases and how to effectively analyze the gathered people has become a hot issue of current research.The method of analyzing crowd behavior and making corresponding decisions by watching surveillance video manually often lacks timeliness and accuracy.Traditional crowd analysis methods,which extract pedestrian feature by different feature extraction algorithms,often have low accuracy and robustness because the features extracted by manual are poorly versatile and cannot accurately reflect the characteristics of dense crowd.In recent years,deep learning has made unprecedented achievements in various fields,and some crowd image analysis methods based on deep learning have been proposed.However,these methods often have huge parameters,so it is difficult to deploy the task of dense crowd analysis in situations of insufficient computing resources such as mobile terminal and embedded device.In addition,the accuracy and robustness of these methods need to be further improved.In view of the shortcomings of the existing methods,a lightweight convolutional neural network and a feature sharing convolutional neural network are proposed for dense population image analysis.The lightweight dense crowd image analysis network proposed in this thesis only contains less than one million parameters,and can analyze the dense crowd image with resolution of 768×1024 at an average speed of 2.5 seconds per frame on CPU.The lightweight network uses a scale-aware module to extract pedestrian features of different sizes,and then regresses these features to a rough density map,the density map is refined and purified by using an auto-encoder to obtain a more accurate population distribution density map finally.The lightweight network proposed in this thesis achieves good performance on two common available datasets,achieves a balance between accuracy and complexity,and is suitable for dense crowd image analysis on devices with insufficient computing resources such as mobile terminals and embedded devices.In order to further improve the accuracy of dense crowd image analysis methods,this thesis proposes a feature sharing convolutional neural network from the perspective of improving feature utilization.In this network,the first ten layers of VGG16 convolution network which is pre-trained by on the ImageNet dataset,are used to extract the universal features,then a feature sharing module S-Module is used to combine the universal features and abstract them into advanced pedestrian features,which makes full use of the features of different levels of abstraction for regressing to a population distribution density map.The results show that the proposed feature sharing network has high accuracy and robustness,and can be used in dense crowd image analysis tasks with high accuracy requirements.