Research On Density Estimation And Crowd Counting Algorithms Based On Convolutional Neural Network

Owing to the urgent practical demands and wide applications for density estimation and crowd counting technology in public safety,intelligent traffic,and urban planning.Density estimation and crowd counting for highly dense scenarios have always been a very hot topic in the field of computer vision.With the rapid development of national economy and the continuous promotion of population urbanization,the urban population is increasing day by day.In addition,with the improvement of people's living standards and the needs of spiritual and cultural,traffic congestion,stampede,terrorist incidents appear in various scenic spots,shopping malls,music and religious gatherings and other scene crowds during the holidays.Which posed a great threat to the people's personal and property security.Relevant departments are continually increasing manpower to maintain public order and ensure public safety,but there are still limitations in monitoring and control by manpower alone.If we can make full use of density estimation and crowd counting technology combined with video surveillance,on the one hand,we can save lots of manual labour and material resources,on the other hand,we can analyze,prevent and deal with these emergencies more effective and faster.In the past few years,due to the rapid development of deep learning,convolutional neural networks have achieved remarkable results in recognition,detection,tracking,person re-id and other hot areas of computer vision.At the same time,the crowd counting algorithm has successfully transited from the early traditional methods to the deep learning methods.Traditional methods are usually limited to the situation of low density of crowd and little variations of target scale.The correlation density estimation methods based on convolutional neural network has been developed to cater for changes including different density levels,scale,perspective distortions,target occlusion and other scenarios.In this paper,we firstly classifies and analyses the traditional crowd counting algorithms according to its characteristics,and points out its shortcomings.Secondly,some classical density estimation networks based on convolutional neural networks are introduced in detail.Subsequently,we proposed some optimized algorithms.The feasibility and validity of the algorithms are demonstrated on some challenging datasets such as Shanghaitech and UCF_CC₅0.The specific works of the paper are as follows:1.Aiming at the drawbacks existing in the FCN algorithm of density estimation and crowd counting based on CNN,this paper proposed an algorithm based on convolutional features fusion.The main idea is:considering that each convolutional layer of CNN contains different hierarchical information.Some feature maps generated in different convolution layers of FCN are effectively fused to obtain more rich hierarchical information and more discriminative features of the target scene.Extensive experimental results have demonstrated the effectiveness and robustness of proposed method.2.Aiming at the shortcomings existing in the MCNN network of density estimation and crowd counting based on CNN,this paper proposed a sequential multi-stage crowd counting network.The main idea is:on the basis of MCNN network coping with scale variations,in order to obtain more discriminative feature information,we try to explore a deeper network architecture.In addition,the vanishing gradient phenomenon may easily appear in the process of back propagation as the network becoming deeper,so we adopt a learning objective function that implements intermediate supervision in each stage.That is to calculate the loss of output in each stage,so as to ensure the normal parameters updating in the training process,and solving the problem that the deep network hard to be optimized.