Research And Application Of Crowd Counting Algorithm Based On Convolutional Neural Network

With the increasing mobility and concentration of population in modern society,the traffic statistics of airports,stations,shopping malls and other places have drawn a lot of attention.The stampede incident happened in Shanghai on December 31,2014 once again sounded the alarm of effective supervision of intensive population,and also indicating that traditional manpower supervision could not effectively cope with the increasing density of crowds.So,a safety management system with high intelligence is badly needed to improve early warning capabilities.To this end,the computer visionbased crowd counting algorithm has been extensively studied,and the related results can not only effectively improve the safety and prevention capabilities of various crowded places,but also bring unlimited business opportunities to various business fields.In recent years,with the rapid development of deep learning technology,more and more scholars have devoted themselves to the research work in the field of crowd counting.The excellent performance of convolutional neural network in image feature extraction and model generalization can solve the problem of counting in high-density crowd images in complex scenes.In order to further improve the accuracy and stability of the crowd counting method,this paper studies the current crowd counting method based on convolutional neural network and proposes a more effective algorithm model,and also designs the corresponding crowd counting prototype system to verify its effectiveness.The specific works are as follows:The crowd counting algorithm based on multi-column convolution neural network(MCNN)and switch convolution neural network(Switch-CNN)are implemented.The design principles,network structure,implementation methods and training strategies of the two algorithms are also analyzed in depth.The performance of the two algorithms was tested and analyzed using the general crowd counting data set.By optimizing the design training strategy,the experimental results close to or better than the original algorithm were obtained.Finally,we lays the foundation for the research of the subsequent higher performance crowd counting algorithm through indepth analysis of various factors affecting the performance of the algorithm.A crowd counting algorithm based on multi-scale multi-level feature fusion is designed and implemented.The common convolution and hole convolution are used to extract the features of the image in different receptive fields.The features in different levels are first down-sampled and then down-sampled in the U-shaped network to generate a higher crowd density map through the fusion of feature maps of different scales and different level to realize the crowd counting.The performance of the algorithm is tested by three typical crowd counting datasets.The results show that the algorithm can accurately count all kinds of dense populations and achieve better results than most mainstream algorithms.Based on the proposed crowd counting algorithm,a crowd counting prototype system based on video surveillance was designed and developed.The system was analyzed and designed from the practical application requirements of crowd statistics in public places such as canteens and shopping malls.The design and implementation of the network communication module,the video processing module and the visual interface are also completed.By generating a density map for the acquired video frames by calling the crowd counting module,the statistics and analysis of the crowd flow are realized which validates the effectiveness of the proposed algorithm and also lays a foundation for the development of the actual crowd counting application system.