Research On Adaptive Crowd Counting Integrating Density Attention Mechanism

Today,with the prosperity and development of the economy,people's travel and social activities are becoming more frequent.Crowds often appear in public places such as train stations,bus stations,subways,shopping malls,and parks.On the one hand,crowding will cause inconvenience to people's travel and activities.On the other hand,if it is not evacuated as soon as possible,it may easily lead to safety accidents.Monitoring these public places and controlling crowd distribution information in real time is one of the effective measures to ensure public safety.At present,video surveillance systems are widely deployed in many public places.Using computer vision-based analysis technology can automatically obtain the distribution information of the crowd.In recent years,deep learning methods have achieved important results in the field of computer vision.Benefiting from this,crowd analysis algorithms based on convolutional neural networks are constantly emerging and have made some progress.However,since deep learning methods are mainly driven by data.When there are different crowd density distribution patterns in the scene,convolutional neural network crowd counting models are easy to overestimate or underestimate the number of people in different density areas.This leads to a reduction in the accuracy of the overall crowd count.In order to overcome this problem,this paper proposes an effective method to alleviate the counting error of different density areas,thereby improving the counting accuracy.The work of this article mainly includes three parts.(1)A Density Aware Network(DANet)is proposed.DANet gives regional masks corresponding to different density levels by learning the crowd distribution in complex scenes.On this basis,the error distribution law of convolutional neural network counting model is analyzed.It provides new ideas for designing a method that can effectively reduce the counting error of people in different regions.(2)An Attention Scaling Network(ASNet)is proposed.ASNet first generates multiple intermediate density maps and scaling factors.Then use the scaling factors and the multiple density level masks provided by DANet to adjust and generate the area density maps of the corresponding areas.Finally,combined with the area density maps to generate the final crowd density map.Thereby it reduces the counting error of different density areas.(3)An Adaptive Pyramid Loss(APLoss)is proposed.APLoss can adaptively divide the scene into pyramid-shaped sub-regions according to the density distribution of the scene.Then it calculates the relative prediction error of each sub-region separately.Finally,the errors of all subregions are summed to obtain the overall density prediction loss value.This loss calculation method can significantly reduce the adverse effects caused by the uneven data during the training process,and improve the generalization ability of the convolutional neural network counting model.Experiments on four challenging datasets prove the superiority of the method proposed in this paper.