Research Of Crowd Counting And Person Search In Monitoring Scene

With development of economic society and people's living standard,people pay more attention to public security issues.Today,intelligent monitoring systems are widely used in public places such as hospitals,schools,shopping malls and railway stations.These monitoring videos can help manager timely solve the problems of public facility and better protect people's life and property,and that can make better use of public resources.Pedestrian search and crowd counting are the hot spots,difficulties and new bright spots in the field of intelligent video surveillance and public safety.It is also a research focus and hospital in the field of computer vision and pattern recognition.Pedestrian search and crowd counting are also of important research and practical value.They can help public safety departments monitor public places,search for targets and warn of danger.In recent years,with the increasing attention of researchers for deep learning,deep learning has also achieved good results in many research fields.Many researchers began to use deep learning methods to solve the problems in their respective fields.This thesis first introduces the development process of deep learning,and the application model of deep learning in many fields and problems.Then this thesis shows the current research status of crowd counting,pedestrian search and pedestrian detection algorithms at home and abroad,and carefully analyzes the characteristics of these methods,and their advantages and disadvantages.Finally,this thesis proposes our own deep learning models for crowd counting and pedestrian search:pedestrian statistics based on deep learning and pedestrian search networks based on deep learning.The proposed crowd counting network based on deep learning,firstly use the pre-trained convolutional neural network to extract the low-level features of the whole image,and then use the statistical analysis method to obtain its statistical features.Finally,this model uses support vector machines to establish a mapping relationship between the statistical characteristics and the number of image pedestrians to directly obtain the statistical results of the number of pedestrians in the image.Our contribution to this model is to use the first layer of neural network features,reducing the complexity of the model.At the same time,statistical analysis is introduced into this model to make statistical analysis of the extracted low-level features.Finally the statistical features are used for regression analysis to get the result of crowd counting.Crowd counting based on deep learning is evaluated and examined the performance in public dataset UCSD,and the absolute and relative errors of the proposed model are all less than those of other algorithms.Therefore,the validity of this model is verified.Our proposed pedestrian search network based on depth learning,firstly uses the ResNet50 residual network to extract the features of the whole image,and then uses the pedestrian's individual labels in the sub-network to find the candidates.Then this network extracts the features in these candidate boxes,and inputs these features into the network of pedestrian detection and pedestrian re-identification respectively to find out where target persons are and who target persons are in the picture.For the pedestrian re-identification network,the network uses multitask learning to learn better pedestrian features representation.In this network,the main contribution is to use the pedestrian's individual labels in the network where the candidate box is extracted,so that the resulting feature representation improves the re-identification of the network while maintaining the accuracy of the detection network.And this network uses multitask learning in the pedestrian re-identification network to enhance the performance of re-identification network.To test the effectiveness of our network,the proposed network is evaluated on the public dataset CUHK-SYSU.The experiment results show that proposed network defeats all the contrast algorithms and models.So the deep learning network proposed in this thesis is effective.