| With the rapid increase of the urban population,there are more and more public safety incidents due to large-scale crowd gathering in recent years.Therefore,real-time monitoring of the number of people is of great significance in ensuring public safety.In the past,manual monitoring of video was not only prone to fatigue but also inaccurate to estimate the number of people due to human subjectivity.With the development of computer vision technology,it has became possible to automatically obtain the number of people and the distribution of people from surveillance videos.In addition,the monitoring of the number of people has great research value in commercial,public transportation and city management.The biggest problem with crowd counting is that the crowd environment is complex and changeable,and people in the crowd have multiple perspectives,low resolution and severe occlusion.In view of the above problems,a multi-scale feature fusion adversarial neural network method is presented.The network uses different sizes of convolution kernels and a pooling filter to extract multi-scale shallow crowd information,then the residual connection fuses deep and shallow crowd characteristics to improves the network's ability of detect small targets.The network train through adversarial methods to guide the model to generate the high-quality crowd density map.The crowd count is obtained by summing the pixel values of the crowd density map.This article conducted a series of training and testing on ShanghaiTech and UCFCC50 which are two major crowd counting data sets,First,four experiments are designed on ShanghaiTech to verify the improvement of crowd counting results by using multi-scale feature fusion,deep-level feature fusion and adversarial loss function.Next,the results of the experiments on ShanghaiTech and UCFCC50 are compared with the results of the current mainstream algorithms,which proves that the method has high accuracy and robustness in complex crowd environments. |
PDF Full Text Request