Multi-Scale Crowd Counting Under Complex Scenes Based On Generative Adversarial Networks

With the advancement of global urbanization,intelligent monitoring has gradually become a research hotspot in the field of computer vision.As one of the core issues of intelligent monitoring,the crowd counting problem is of great significance in application scenarios such as crowd current limiting drainage.At present,the research work on population counting has made great progress.However,in different scenarios,research on solving the problem of inconsistent image size of the crowd still has great challenges.In recent years,deep convolutional neural networks have achieved outstanding results in the field of computer vision research.Its outstanding performance in image feature and model generalization has effectively solved the feature extraction problem of crowd counting under complex background.In order to extract the features related to scale,the current crowd counting model based on convolutional neural networks adopts the structure of multiple columns or multiple networks,but such structures have problems such as large amount of training model parameters.In order to effectively solve the problem of extracting scale-related features in population counting,this paper proposes a multi-level convolutional neural network structure,which extracts multi-scale features in the network by adding multi-scale convolution modules on the single-column convolutional neural network.At the same time,in order to solve the problem of small target detection in high-density crowd images,the algorithm uses a multi-level convolutional neural network to fuse the features of global features and local features at different scales.Since the refraction core of the low-level convolution layer has little receptiveness,the features of the low-level convolutional layer contain more characteristic information of small target objects.Multi-level convolutional neural networks contain low-level convolutional layer detail features and high-level semantic features.Enhance the network's perception and technical ability to small targets by incorporating low-level details into high-level semantic features.At the same time,in high-density and complex crowd scenarios,there is a high degree of similarity between pedestrians and backgrounds,resulting in poor quality of population density maps predicted by neural networks.In order to improve the quality of predicted density maps,this paper proposes a multi-scale based on multi-scale.A population counting method for generating a confrontation network?MS-GAN?is introduced,and a discriminator is introduced as a supervisor for generating a density map,and the purpose is to guide the generation network to predict a high quality density map.The use of confrontational training methods enables convolutional neural networks to generate higher quality density maps.In this paper,the crowd image and the crowd density map are superimposed into the discriminator,and the discriminator is trained to discriminate the density map of the real label and the predicted low-quality density map.In the process of predicting the network of the training density map,the anti-loss is added to improve the quality of the generated density map.Experiments on UCF_CC₅0,Shanghaitech,UCSD and Mall four population count datasets show that MS-GAN is superior to MCNN in accuracy and stability,which significantly improves the quality of the predicted density map,thus improving the accuracy of crowd counting.