Research On The Number Of People Statistics Method For Non-uniform Population Distribution Scenes

Crowd counting plays a more and more important role in public safety?activity planning and space design.The crowd counting method based on computer vision has received more and more attention as a kind of crowd counting solution that is contactless and does not require the active cooperation of monitored objects.The traditional crowd counting method can well calculate the crowd number in sparse crowd scenes,but it is not suitable for crowded scenes with dense people,because there are many interference factors in crowded scenes,such as serious occlusion and background clutter.The crowd counting method based on deep learning can well calculate the crowd number in complex scenes,but the crowd counting method still faces many challenges.Firstly,almost all the crowd counting methods based on deep learning are aimed at the situation that the crowd is almost full of the whole scene.However,the images obtained by the video surveillance and aerial photography are not necessarily shooting at the crowd.More often,the crowd only occupies a part of the whole image.If the crowd counting is carried out directly on the images containing a large number of cluttered backgrounds,it is very easy to affect the counting results.Secondly,there is perspective distortion in the same scene.When facing crowd scale change scenes,for example,the head near the camera is large,the head far away from the camera is small,and the crowd scale changes greatly in different scale crowd scenes.This requires dealing with problems of different scales at the same time.In view of the shortcomings of the existing methods,the research contents and main research work of this paper are as follows:(1)We propose a convolutional neural network with background exclusion for crowd counting.Firstly,we divide the image into blocks and then use the residual network to determine whether each block contains crowd,to eliminate the clutter background area and avoid the background interference to crowd counting.Secondly,we use the dilated convolution and asymmetric convolution to estimate the crowd density map of image blocks containing crowd.Finally,the crowd density map of all crowd areas is integrated to obtain the crowd counting results of the whole scene.We collect some images of more general scenes,such as the crowd is only a part of the whole image,and construct non-uniform population distribution(NDC 2020).dataset.In addition,We test the effectiveness of the proposed method on UCF_CC₅0?ShanghaiTech and NDC 2020 dataset.Our method achieves good performance on UCF_CC₅0 and ShanghaiTech_Part_B dataset.Our method is superior to the existing crowd counting methods on NDC 2020 dataset.(2)We propose a crowd counting model based on multi-scale dilated convolution.To further solve the crowd scale change problem based on the convolutional neural network with background exclusion,we propose a multi-scale dilated convolution crowd counting model.Firstly,we preprocess the image in the crowd scene.Secondly,we use the residual network to determine whether each block contains crowd,to eliminate the clutter background area,and select the image block containing crowd.Thirdly,the image blocks containing people are used as the input,and the multi-scale dilated convolution estimate the density map.There are two ways to design the Multi-scale dilated convolution model:the first is Cascaded Hybrid Dilated Convolution network(CHDC),which use the dilated convolution combined with mixed dilated ratio to solve the problem of different crowd scale;the second is Parallel Multi-scale Dilated Convolution network(PMDC),which use three columns of dilated convolution with different dilated ratio to extract different crowd scale,and integrate the output.Finally,the crowd density map of all crowd areas is integrated to obtain the crowd counting results of the whole scene.We test the effectiveness of the proposed method on the existing crowd datasets and NDC 2020 dataset.Our method achieves good performance on ShanghaiTech_Part_B dataset.Compared with the results of convolution neural network with background exclusion,the results of crowd counting based on Multi-scale dilated convolution are further improved on ShanghaiTech_Part_A and UCF_CC₅0 dataset.The method based on Multi-scale dilated convolution is better than the existing methods on NDC 2020 dataset.(3)Design and implement the crowd counting system.Users can use the crowd density distribution and crowd counting functions provided by the crowd counting system to view the crowd regional density distribution and calculate the total crowd number in the image.