Research And Inplementation Of Crowd Counting Method Based On Multi-column And Scale Adaptive Convolution

With the continuous increase of the density of population in big cities,the phenomenon of large crowds gathering is becoming more and more common,and the scale of gathering is also getting larger and larger.Timely and accurate estimation of the number of crowds in various places is not only helpful for the relevant government departments to manage and control crowds gathering,but also can provide tourists with real-time crowd gathering degrees,so that they can make the best travel plan in time.The degree of crowd density and sparseness varies greatly,and in camera equipment,the differences in the head size of the image is huge due to the equipment perspective change causes people head perspective effect.In the existing crowd counting algorithms,the neural network based on detection or density estimation is generally adopted for crowd estimation.However,the single method is difficult to obtain the expected effect when applied to the real application scenarios where the crowd density and head scale difference vary greatly.In view of the above problems,this master's focuses on the research of adaptive crowd counting algorithm,comprehensively uses the neural network based on detection and density estimation to predict the number of crowds,and further improves the crowd density estimation network enhances the generalization performance of the model.The main work is as follows:Firstly,a crowd counting algorithm based on adaptive selection of detection and density estimation is proposed,through dividing an image into nine pictures averagely,using a classifier for the crowd density and head size classification,combined with the classification results to adaptively selected detection module or density estimation module to implement local block counting,effectively improve precision of the crowds counting in the image that the population density varies greatly.Secondly,a crowd counting algorithm based on multi-column and adaptive convolution is proposed.In the density estimation network,an adaptive convolution kernel is proposed to optimize the fixed convolution kernel used in the density estimation network.This method can further improve the counting accuracy of image blocks with large differences in the size of human heads in the high-density population with perspective effect.On the real scene datasets,this paper compares the proposed crowd counting network model algorithm with the crowd counting algorithm network in recent years,the experimental results show that the proposed algorithm in the MAE(Mean Absolute Error,MAE)and RMSE(Root Mean Squared Error,RMSE),compared with other people counting models,for different population density and the head size change has better accuracy and robustness.