Research On Crowd Density Estimation Based On Head Box Prediction

In recent years,as the world population surge,poor crowd management has caused many problems,such as crowd crushes and blockages causing safety accidents,which increasingly requires computing models to analyze highly dense crowd by using video sources from surveillance cameras.Crowd counting is a key component of this automated crowd analysis system.This includes estimating the total number in the population,and the population density distribution throughout the region.In recent years,the success of crowd counting methods can be attributed to the development of deep learning.The crowd counting methods based on density map estimation in deep learning often use the density map generated by Gaussian Blur as the learning target,and convolutional neural networks(CNN)are used as frameworks,then use the pixel level mean square error(MSE)loss function for optimization.However,the head dot annotations exist random variances inside the head area,not all the label points in the positive middle of the head area.So the generated density maps use to ignore the information of the head size and shape.When using the density map as the learning target,we do not require the loss function of the mean square error(MSE)which is pixel-level and such strict,instead the loss function based on retaining the sign of the head area which called Sign-Loss is proposed to train the crowd counting network in this paper.Computing Sign-Loss is based on the head box area,however most of the existing crowd counting datasets only have dot annotations of head positions.So this paper propose head box prediction network(HBPNet)to estimate its corresponding head size for annotation dot.Due to the human head size differ in the crowd image,this paper propose a multi-scale crowd counting network Sign-Crowd Net based on high-resolution network(HRNet)to verify the effect of the proposed loss function based on retaining the sign of the human head area.Furthermore,in the visualization results of density map estimated by crowd counting network,there will be some false response problems in the background area.Sometimes,even if the total number of crowd estimated is close to the real number,some of the estimated values may come from the response of the background area.Therefore,attention mechanism is added to the head region to optimize the model.The experimental results on multiple datasets show that the performance of the model with head region attention mechanism is improved.In summary,this paper analyzes the problems of dot annotation variance,scale variation and background error response in crowd counting,and puts forward the corresponding solutions.The experimental results on the mainstream datasets of the crowd counting show that the proposed method has a significant improvement in the evaluation index.