Research On Crowd Counting Method Based On Convolutional Neural Network And System Implementation

Crowd counting is an important research content in the field of computer vision,and it can be applied to people counting in squares,stations,concerts,schools and other crowd gathering places.At present,the deep learning method is the mainstream method in the field of crowd counting,it all thanks to the development of hardware equipment and the improvement of deep learning methods.However,there are still many factors that affect the performance of crowd counting in practical applications.Such as,the change of the head scale will affect the counting performance,the uneven distribution of the crowd will affect the extraction of head features,and the background interference will cause misjudgment.The work of this paper will focus on these questions.Crowd counting on the basis of density map is the main method at present.In order to obtain a density map that is closer to the real scene,this paper uses the geometric adaptive Gaussian kernel method to generate the true value density map,and different β values should be used for different density scenes,so that the reference density map can be better correspond to crowd images of different viewing angles and different degrees of density,which helps the network to predict a more accurate density map.Aiming at the change of head scale,background interference,and uneven distribution of crowd in crowd images,an encoding and decoding multi-path network integrating attention mechanism is proposed.Extract multi-scale features using the inter-level structure of convolutional neural networks and improve head pixel attention using the encoder-decoder structure.We further propose a multi-scale attention crowd counting model fused with dilated convolutions,the front-end extracts five feature levels from VGG19 to solve the problem of large-scale changes in crowd counting.It adopts a dual-column branch structure and uses attention map branches to reduce background interference.The density map branch of the column network structure extracts contextual information and enhances crowd feature learning.And it use the dilated convolution structure with different expansion coefficients to deepen the network depth while ensuring the receptive field,strengthen the multi-scale learning ability of the network,and combine the attention map to filter the background interference to strengthen head attention.In the multi-scale attention crowd counting model fused with dilated convolution,Euclidean loss and attention mapping loss are combined to speed up the network convergence speed and give more attention to the head pixels in the density map,which is conducive to the regression of higher quality density graph.The crowd counting management system is designed and developed by comprehensively applying the research results,which can realize the density map prediction of a single crowd image and display the counting results.