Research On Crowd Counting Based On Convolutional Neural Network

In recent years,stampede accidents have occurred frequently all over the world,especially in populous countries.Most of these accidents took place in urban public places where people gathered for celebrating festivals,religious activities,sports events and so on.Accurate crowd counting in surveillance videos provides information support for constructing the new smart city which treats video and image processing and analysis in public safety as the core.It plays an important role in enhancing urban management and preventing public crises.Besides,there are many migratory applications of crowd counting in other fields,such as vehicle counting in transportation,cell counting in medicine and species counting in the wild.So it also has a positive impact on all aspects of society.The purpose of crowd counting is to count the number of people or estimate the density of large-scale population in a given scene.One of the difficulties is that individual size of the crowd varies greatly.Convolutional neural network(CNN)has played an important position in the field of computer vision due to its excellent results in video and picture processing tasks.Thanks to the advantages in feature learning and performance,it is and has become a trend to use CNN to study the subject of crowd counting.In this paper,we focus on the multi-scale problem of crowd individual,and the specific work is as follows:First of all,we design a series of shallow convolutional neural networks(SL-Net)to compare the sensory ability of different size of receptive fields in the CNN structure on different size of individuals by using error indexes on the simulation dataset,Mall dataset and library visual angle dataset respectively.It was concluded that larger kernel size is suitable for perceiving small-scale objects with global feature information and smaller kernel size is suitable for perceiving large-scale objects with local feature information.Secondly,we propose a new type of single-column multi-scale perceptual convolutional neural network(SCMS-Net)based on the network architecture of VGGNet8 which can extract scale-related features for learning by multi-scale perception module.Other improvements include the multi-scale augmentation of the Mall dataset to further enhance the adaptability of the network to multi-scale targets,using global average pooling instead of the fully connected layer to reduce the amount of parameters and introducing batch normalization into many parts of the network to overcome training difficulties and slow convergence caused by deeper layers.In addition,we also propose the improved network SCMS-Net++,which not only uses the residual structure to achieve equivalent fusion of multi-scale features in order to improve sensory ability of the network to multi-scale targets but also implements multi-size input.Finally,we verify the effectiveness of the designed and improved networks and compare the corresponding results in each stage of the improvement.The final networks,SCMS-Net and SCMS-Net++,were proved to be superior to most of the existing crowd counting networks tested on Mall dataset in terms of accuracy and stability.Compared with the mainstream complex multi-column network structure and multi-network hybrid structure,the singlecolumn network structure in this paper is easier to be trained and optimized,and saves training time and computing resources.It has great practical significance and application value.In conclusion,we propose a type of single-column crowd counting CNN to realize multiscale features extraction and equivalent fusion from the horizontal structure,which provides a new idea for designing multi-scale convolutional neural network.