Crowd Counting And Crowd Abnorma Behavor Detection Based On Generation Adversarial Networks

With the development of image processing and video analysis technology,video surveillance has been widely used in social management,public security prevention,serving people’s livelihood and other aspects.Therefore,it is of great theoretical significance and application value to conduct crowd behavior analysis from crowd counting and crowd abnormal behavior detection.This paper adopts the method based on MS-GAN to realize crowd counting and the method based on MFF-GAN to realize the detection of abnormal crowd behavior.The main work and conclusions of the research are as follows:1.A crowd counting algorithm based on MS-GAN is proposed.The first ten layers of the improved VGG16 are used as the front-end backbone network of the generator.In order to obtain the multi-scale context information of the crowd,the back-end introduces atrous convolution with expansion factors of 1,2,3,and 4 for multi-scale information aggregation to solve the problem of multi-scale changes.The discriminator network Patch GAN is trained against the generator network,guiding it to generate high-quality local count maps.The experimental results show that the proposed MS-GAN algorithm is better than most mainstream algorithms.Compared with the CSRNet network,MAE values of MS-GAN tested on Shanghai Tech dataset are reduced by 6.1 and 3.8,and MSE values are reduced by 16.9 and 5.2,respectively,which can predict the number and distribution of the crowd well,with high accuracy and robustness.2.A crowd abnormal behavior detection algorithm based on MFF-GAN is proposed.The generator front-end encoder network introduces 3D convolution,ATT-LSTM for extraction and temporal prediction of important frames of the input video sequence.The prediction branch of the backend decoder,which is used to predict the next frame of the input video sequence.The discriminator network SN-Patch GAN is trained against the generator network,prompting it to generate sharper images and more accurate predictions.Introduce Flow Net2.0 to extract optical flow weights to estimate crowd appearance and overall motion trajectory features.The experimental results show that compared with the algorithm of Gao et al,the proposed MFF-GAN algorithm can improve the AUC value of the CUHK Avenue dataset by 1.4%,and the AUC value of the USCD Ped2 dataset by 0.1%,which can improve the detection performance and realize abnormal behavior real-time detection and localization.