Dense Crowd Counting Algorithm Based On Multi-scale Attention

At present,with the increasing number of large-scale public events,more and more people are gathered,such as celebrations,concerts,sports events,public parades,etc.Therefore,effective supervision of dense populations is necessary.In the field of computer vision,crowd counting can provide technical support for the supervision of large-scale crowd gathering,and its related research is also very active.Most of the early algorithms used traditional detection and regression methods,and it was difficult to effectively solve problems caused by factors such as occlusion,perspective distortion,proportional changes,and diverse population distribution.With the increase of crowd density,occlusion among people becomes more a-nd more serious,coupled with influence of actors such as uneven distribution and light,which have which have challenged the counting of dense crowds.In recent years,more people have adopted deep learning methods to achieve population counting.These methods use a convoltt convolutional neural network to generate a density map,and then sum the den-sity maps to count the number of people.Compared with traditional methods,this meth-od has better accuracy and applicability.This paper proposes a dense crowd counting algorithm based on multi-scale attention.The algorithm consists of three parts: feature extraction,multi-scale attention density map generation and density map fusion.Our backbone network is composed of the first five layers of vgg16,and the multi-scale attention module and density map generation module are added to the feature extraction module separately.The fusion module connects the multi-scale attention module and the density map generation module.Specifically,the image features are first extracted through the backbone network,then the features of different scales are separated by the attention model,and finally the reference density map and the scale attention density map are fused.Among them,scale attention includes three parts: attention prediction,baseline density map prediction and Inception structure.Multi-scale attention utilize.Convolution kernels of different scales of the Inception structure to achieve feature separation and improve network performance.We performed experiments on three mainstream crowd counting datasets,and experiments have shown that our method has more effectiveness.