Research On Crowd Density Estimation Method Based On Attention Mechanism

When the crowd density is too high,it is easy to cause accidents and accelerate the spread of infectious diseases.At present,the research on crowd density is mainly divided into traditional methods and deep learning methods.Traditional methods use artificial feature extraction,which consumes a lot of time and energy,and often fails to respond to abnormal phenomena;while the deep learning method uses the CNN network to learn and train the content of the crowd image.There is no need to extract features manually,the CNN network is more intelligent,and its cost and performance are better than traditional methods.Based on the current situation of the CNN model,this thesis introduces the attention mechanism into this type of model,and proposes a CNN model based on multi perception feature extraction and fusion.Then,this thesis optimizes the proposed model to improve the overall performance of the model.This thesis mainly completes the following tasks:(1)The CNN model is integrated with the attention mechanism to build a multi-sensory feature extraction CNN model.First,according to the current research status of crowd density and the size of heads,a three-column CNN model is built,and different branch CNNs use filters of different sizes to achieve head size adaptation.Then,based on the three-column CNN model skeleton,the CBAM module in the attention mechanism is introduced into the CNN model.According to the characteristics of CBAM,it is set at the first and last convolution of each column of CNN,so that the model discards invalid information and focuses effective information.(2)Optimize the initial model(the multi-sensory feature extraction CNN model).First,based on the feature extraction and fusion,this thesis adopts multi-scale and cross-level optimization,the model at the shallow level uses the Inception structure for multiple feature extraction and fusion,and the deep-level model realizes cross-level feature fusion through the ShortCut structure.Then,based on the characteristics of crowd density estimation,the model needs to have lightweight features,so 1*1 convolutional layer is used.Not only reduces the amount of model parameters,but also allows input of crowd images of any size.Finally,the model is optimized by introducing BN,the BN can make the model better fit and better convergethe by making the data distributed in a certain fixed area.(3)Build the model designed by this thesis through Tensorflow and Keras framework.Based on the Crowded Data Set made by this thesis to achieve model training,adjustment and testing.Then,based on the dataset produced in this thesis,the model proposed in this thesis is compared with the classic similar models Cifar10 and MCNN.According to the loss and accuracy value change curve,it shows the model proposed and optimized in this thesis has a faster convergence speed and better accuracy.