Research On Crowd Counting Based On Multi-scale Dilated Convolutional Neural Network

With the obvious growth of the urban population and the improvement of people's material living standards,various large-scale population gathering activities continue to increase,which can easily lead to trampling accidents.Intelligent real-time monitoring of crowd scenes has attracted people's significant attention.Thus the automatic recognition of real-time monitoring videos and analysis of the crowd density to take instant precaution is not only of great significance to the protection of social and public safety,but also plays a significant role in smart urban planning,public security investigation,and transportation regulatory and other fields.In recent years,the crowd counting algorithm has become a research hotspot in the field of computer vision at home and abroad.However,due to the problems of different crowd scenes,severe occlusion,background confusion,and perspective effects,it has brought great challenges to the research.Traditional crowd counting algorithms are usually implemented algorithms based on detection or regression,which tends to be only applicable to low-density crowd scenarios.With the emergence of technologies such as artificial intelligence and convolutional neural networks,it has brought a new chance for the crowd counting problem,but most existing algorithms are still only applicable to specific scenarios and have difficulties in adapting to the changes of people's scale.This paper proposes a multi-scale dilated convolutional neural network(MSD-CNN)for crowd counting,the main innovations of which are summarized as follows:1.Use the density-dependent geometric adaptive Gaussian kernel.Taking into consideration the problem that geometric adaptive Gaussian kernels are likely to cause large errors in sparse crowd scenes,this paper proposes to select different ? coefficient values according to the crowd density levels in the scene,so as to adjust the head size automatically to adapt to crowd scenes with different density and generate more more realistic reference density map.2.Design a multi-scale feature fusion module is designed in the single-column network structure.Since the multi-column or multi-input network structure has the disadvantages of complicated design and large amount of calculation,in order to reduce the complexity of the network and adapt to the changes of people's scale in the image at the same time,this paper designs a single-column network structure to extract feature.In reference to the Inception module and VGG network idea,a multi-scale feature fusion module is designed in MSD-CNN to aggregate crowd features of different sizes.3.Introduce dilated convolution to take place of the pooling operation.Because the pooling layer will reduce the image resolution and lose important detail features of the original image,MSD-CNN replaces the pooling layer with dilated convolution layers with different dilation coefficients to increase the receptive field while maintaining the image resolution.Besides,it's beneficial to prevent the loss of small-scale features and improve the accuracy of counting results in dense crowd scenarios.This paper validates the MSD-CNN on two espicially representative public crowd data sets.Experimental results indicate that the MSD-CNN crowd counting model proposed in this paper can reduce the complexity of the network and has better adaptability to different crowd scenarios at the same time,with which we can estimate the population density distribution and the crowd count more accurately.Compared with several mainstream crowd counting algorithms,this method has higher accuracy and better robustness.