Research On Crowd Counting Algorithm Based On Light-weight Convolutional Neural Network

In public places,too many people are likely to cause stampede events and easy to spread diseases,and then endanger public safety,causing major violations of personal and property security.Crowd counting algorithm based on deep learning can avoid manual monitoring and efficiently control the crowd,which has become a hot research direction in recent years.However,for the sake of precision transcendence,most current crowd counting algorithms often have complex training strategies,complex model structure,too many parameters and computational complexity.Based on this,this paper studies the light-weight crowd counting algorithm,through dividing the algorithm model into coding structure and decoding structure,build a multi-scale,cross-level connectivity light-weight crowd counting algorithm,the main work is as follows:Firstly,based on the basic structure of MobileNetV2,a light-weight coding structure is constructed,and the decoding structure is improved to build a light-weight crowd counting model called MobileFusion.Aiming at the multi-scale problem of the crowd in the image,an improved feature pyramid structure is introduced into the decoding structure,and a light-weight improved dense up-sampling convolution module is introduced into the decoding structure.Experimental results show that the proposed decoding structure can effectively suppress the checkerboard effect,and the output density map has more detailed information.In addition,MobileFusion shows excellent performance when tested on ShanghaiTech,UCF-QNRF and UCF CC 50 datasets.Secondly,for coding structure,MobileFusion was improved to build a more light-weight crowd counting model called GhostFusion.The coding structure of the model is based on the Ghostmodule in GhostNet.In addition,in order to improve the precision of the model,the light-weight spatial attention mechanism and channel attention mechanism are introduced into the coding structure and the effectiveness of adding these attention mechanisms to the model is verified by experimental comparison.Among them,the light-weight spatial attention mechanism achieves the highest accuracy without significantly increasing the model parameters and computational load;Finally,it is difficult to train GhostFusion effectively and the model precision is low.In this paper,the original Euclidean distance is improved,and the count loss is introduced into multi-task learning.Experimental results show that the improved loss and multi-task loss can improve the accuracy of the model.In addition,GhostFusion performs better on the Shanghi Tech and UCF-QNRF datasets than on other mainstream models of equal complexity.