Research On Crowd Counting Algorithm In Complex Scenarios

With the rapid development of society,the increase in large-scale gatherings and parades,and the frequent trampling incidents in the activities caused a large number of casualties.Therefore,the study of crowd counting has become one of the main research directions in the field of computer vision in recent years.In this paper,through in-depth analysis of the advantages of convolutional neural networks,full consideration of the internal connections between the feature layers,and from the perspective of random grouping distribution,the following three different population counting network structure models are proposed.(1)A Feature Self-Learning Multi-Scale Consensus Generative Adversarial Network(FSLGAN)model is proposed.This paper analyzes in detail the influence of different parts of the FSLGAN network model on the results of crowd counting.The method first uses an adaptive kernel to generate the true density map of each data set;secondly,in order to adapt to multi-scale,the network is divided into two sub-networks,each extracting different scales Features,using local and global complementary functions to provide more discriminative semantic features;then add a feature self-learning module in the density map generation stage of the two sub-networks to reduce the impact of changes in the image due to lighting,deformation,occlusion and other factors.Improved the model's acquisition of useful features in the density map generation stage;finally,in order to suppress the error between the subnetworks,a multi-scale consistency loss was added.A large number of experiments on several common population data sets show that the FSLGAN proposed in this paper is effective and has good robustness.(2)A Feature Channel Spatial Attention Feature Fusion Network(CSAN)is proposed.Due to the diversity of the population data set,the convolutional neural network has limitations on the features extracted by the same processing method at different scales.This paper proposes a new CSAN model for dense scenes.The network is mainly composed of three parts.First,the first 10 layers of the VGG network are used as the front-end network.The second part of the network is composed of the channel space attention module,where the channel attention is responsible for extracting useful features between the channels of the feature layer.The spatial attention module is responsible for feature extraction on the channel space.The third part of the network uses 6-layer convolution to increase the receptive field of the network.The experimental results in different data sets prove the effectiveness of the CSAN proposed in this paper,and it has excellent performance.(3)A hybrid loss feature channel spatial attention depth feature fusion network(HLCSADN)is proposed.Due to the camera shooting angle,the population distribution in many pictures is extremely uneven.Many methods try to solve this problem by using multicolumn or multi-branch networks.However,due to the limitation of the number of columns or the number of branches,the extracted features are not uniform which cannot express the ability of outliers in the crowd in a picture.This paper proposes that the network is based on CSAN.The network adds migration learning parameters in the early stage of the network and uses a new hybrid loss in the network training stage.The loss function is composed of mean square error loss and absolute error loss.Loss can effectively solve the problem of outliers.Experimental results on different data sets show the effectiveness of the HLCSAN proposed in this paper.To sum up,this paper proposes network models based on deep learning: FSLGAN,CSAN,and HLCSAN,and the methods proposed in different chapters have a certain improvement.Compared with FSLGAN,CSAN has not only achieved greater accuracy improved and greatly shortened the training cycle,HLCSAN has achieved the lowest calculation error compared to CSAN and FSLGAN,and a large number of experiments have proved that they have practical application value.