The Research And Application Of Crowd Counting Method Based On Deep Learning

Intelligent video surveillance system is an important technical support to ensure national safety,and its application scenarios in real life are widely distributed.The task of crowd counting is one of the important tasks in the intelligent video surveillance system.With the growth of people's awareness of safety prevention and control and the improvement of social science and technology,the task of crowd counting method plays an important role in controlling crowd flow in public places,maintaining public security order and controlling novel coronavirus(2019-n Co V).At present,the mainstream crowd counting task is implemented based on the density map estimation method,and its model architecture is constructed by deep learning neural networks.Generating high-quality density maps is one of the main goals of this type of method.However,in the process of neural network training,image upsampling is usually used to assist the output.In the process of upsampling,the phenomenon of loss of effective information will inevitably occur,and the output effect will be damaged.In addition,the current crowd counting models based on deep convolutional neural networks face a common problem,that is,millions of network parameters and huge calculation costs make the training process extremely difficult.Therefore,designing a lightweight crowd counting task model is an important challenge now.In response to the above problems,this article has conducted the following researches:(1)Aiming at the problem of the loss of effective information in the upsampling process of network training,which causes the loss of the output effect,this paper proposes a crowd counting method based on spatial joint upsampling(SJU),which uses VGG-16 as the backbone.The backend of the network adds a cascaded joint upsampling model,which combines multi-layer highresolution feature maps to extract rich pixel details and spatial context information,and finally achieves the effect of reducing computational complexity and improving accuracy.The MAE/MSE of this method on the dataset of Shanghai Tech A and B are 62.1/99.8 and 7.6/11.5,respectively,and the MAE/MSE achieves 61.3/99.2 after adopting the curriculum learning mechanism.In addition,this method has also achieved excellent results on the large-scale complex dataset NWPU-Crowd,with a MAE/MSE of 105.1/419.3.The experimental results show that the SJU method has better performance than most population counting methods.(2)This paper proposes a model optimization scheme based on knowledge distillation on the basis of the aforementioned spatial joint upsampling crowd counting model.The specific method is to use SJU as a teacher model,and then obtain a lightweight student model S?SJU through the knowledge distillation model compression training strategy.Experiments have proved that the model optimization scheme proposed in this chapter can allow the student network to fully absorb the structured knowledge of the teacher network,so that the compressed model can reduce the computational cost of the model counting task under the premise that the accuracy is sacrificed as little as possible.(3)This article also designs and implements an intelligent crowd flow monitoring system whose purpose is to verify the feasibility and effectiveness of the model and optimization scheme proposed earlier.Firstly,the application significance of the system is discussed.Secondly,the system implementation needs analysis,outline design,detailed design and specific realization are carried out according to the software development process specification.Finally,the partial function effect diagram of the system is shown.Users can upload images of the crowd to be monitored on the front-end browser,or type in valid data such as date data,and the system immediately calls the trained model deployed in-house or calls a database query to return real-time or historical monitoring results.