Research On Crowd Counting Algorithm Based On Deep Learning

With the continuous development and progress of the society,human activities are increasing,and there are more and more gathering places.Accurate estimation of crowd density and quantity can provide very effective information for predicting crowd activities,which is conducive to ensuring social public safety.Crowd counting has become a research hotspot in the field of Computer Vision.Early crowd counting methods were mainly based on detection and regression.In recent years,deep learning has attracted more and more researchers with its outstanding ability in feature learning.In view of the fact that deep learning methods have higher counting accuracy than traditional methods in crowd counting,this paper adopts the methods based on deep learning to achieve dense crowd counting.The main work is described as follows:1)This paper proposes a high-density crowd counting method consisting of a pixellevel attention mechanism and an improved single-column network for crowd density estimation.Aiming at the problem of uneven population distribution and numerous network learning parameters,this method firstly uses full convolutional neural network to classify the crowd image at the pixel level,and then uses the improved single-column convolutional neural network to estimate the crowd density,so as to get the total number of people by regression.The proposed method is tested on three public and popular crowd counting datasets: Shanghai Tech,UCF?CC?50,and World Expo '10.Compared with the current mainstream crowd counting algorithm,the proposed method has better performance in crowd density estimation.2)This paper proposes a crowd counting method based on Generative Adversarial Network structure.This method uses U-net neural network as a generation model,and generates a population density map by encoding-decoding.The method designs a comparison experiment for two input methods of image(original image and image block).In addition,the dilated convolution is introduced into the discriminative model,and experiments are performed on the Shanghai Tech?part A dataset.The results show that the model obtained by using the image block input method and the discriminative network introducing dilated convolution has the best performance compared with the comparison method.