The Research On Static Image Dense Crowd Counting Algorithms Based On Convolutional Neural Network

The computer vision-based crowd counting technology has been gradually applied to many fields such as public safety,and has achieved good results.The needs of different fields bring different challenges to the crowd counting.The crowd counting has been transitioned from the early simple scene to the ability to adapt to complex scene counts.As the application scenario is complicated,the requirements for counting people are getting higher and higher.Therefore,designing,perfecting and improving the crowd counting algorithm has important research significance and application value.The development of deep learning and the excellent performance of convolutional neural networks in image processing have facilitated the application of convolutional neural networks to crowd counting.After carefully studying the typical convolutional neural network crowd counting algorithm,the design idea of integrating CSRNet algorithm and MSCNN algorithm is formed,and a crowd counting algorithm based on convolutional neural network is proposed for crowd counting in highly dense aggregation scenarios.The specific research work includes four aspects.(1)Through the accumulation of relevant literature and theoretical knowledge,the basis of the crowd counting algorithm based on convolutional neural network is known.Carefully read,learn and analyze related literatures and algorithms on convolutional neural networks,traditional crowd counting and crowd counting based on convolutional neural networks,summarize the advantages and disadvantages of existing algorithms,and form improved ideas on this basis.The algorithm of this paper is proposed.(2)Construct a crowd counting model based on convolutional neural networks.In order to effectively extract the characteristics of the population and reduce the crowd counting error,a multi-scale cavity convolutional neural network algorithm model was constructed by using MLP convolutional layer,MSB convolutional layer and cavity convolutional layer.Thedesigned network architecture has different fields of view,and can adaptively extract features of different sizes of the image to ensure high performance of the network model.(3)Image ground truth annotation algorithm design.In the ground truth annotation in this paper,the human head in the crowd is selected as the object of the sample labeling,which can reduce the error caused by the mutual occlusion between the background and the crowd and the crowd.A ground truth annotation is performed using a two-dimensional Gaussian distribution method to generate a high quality density map.(4)Algorithm experimental verification.Using the PartA and PartB training datasets of the ShanghaiTech dataset,train the network model and use the PartA and PartB test datasets of the ShanghaiTech dataset to test the performance of the trained model;use the Mall dataset for the control experiment and verify The mobility of the network model.The network model is verified by multiple angles to prove the feasibility and effectiveness of the algorithm.The innovation follow in:(1)The network structure is highly targeted.In order to better adapt to the crowd counting,the convolutional neural network model was constructed by using the cavity convolution,MLP convolutional layer and MSB convolutional layer as the basic unit.The experiment verified the validity.(2)Adaptability of feature extraction.In this paper,the network can adaptively extract features of different sizes according to the size of each column of the convolution layer in the MSB convolutional layer.The shortcomings are:(1)Due to the constraints of the network structure,the network's learning ability is limited,and the training time is too long,resulting in a decrease in the realtime nature of the network.(2)Some key features are discarded when feature extraction is performed,and the characteristics of the human head are too small,and similar features are likely to occur in the environment,causing errors;therefore,the crowd can be prejudged,whether it is a person,and then feature extraction is performed again.