Multi-Task Density-Aware Learning Neural Network For Crowd Counting

With the gradual growth of China's population base,in public areas such as stations,squares,and parks,the frequency of crowd gathering has become higher and higher,and the density of crowd gathering has become larger and larger.The frequent high-density concentration of the crowd not only brings great security risks,but also hinders the healthy development of the city.Therefore,a set of crowd counting algorithms that can be applied to real scenes is proposed,which is of great significance for public security management,regional spatial planning,and information resource acquisition.In recent years,machine learning and deep learning have developed rapidly,the crowd counting algorithm based on convolutional neural network has shown excellent performance in the field of crowd counting.This paper hopes to combine the computer vision and deep learning to generate a population density map and accurately estimate the number of people.However,due to,for example,occlusion,scenes are disorderly,image resolution is too low,illumination is uneven,and multi-scale of target features interfere with,and it is difficult to produce high-precision detection results.The most advanced methods are usually based on multi-strategy convolutional neural networks,or multi-column convolutional neural networks to learn the characteristics of targets.The former introduces more parameters in the network,consumes computing resources,and is difficult to be practically applied.The latter has limitations due to structural redundancy and low parameter utilization.To analyze the limitations of the existing methods,this paper first proposes a single-column convolutional neural network BLCNN(Basis Line CNN)for population density maps,which is used for population counting in complex backgrounds.Through a large number of experiments,the BLCNN network has achieved the accuracy of the first-line level at home and abroad by using as few parameters as possible,laying a solid foundation for the next step of integrating density classification information.In order to further improve the accuracy of the results,even in the same scene,the population density distribution still has obvious changes,resulting in the problem that the generated density map error is high.In this paper,Multi-Aware CNN(Multi-Task Density-Aware CNN),a multi-task learning convolutional neural network,is proposed.The density-aware branch is used to fuse the level information of the population density in the image into the whole network.Improve the detection accuracy of the network.This paper also uses the joint training strategy to select the BLCNN network as the pretraining network of the Multi-Aware CNN network to participate in joint training,speed up the model convergence and improve the training efficiency.In addition,in order to enrich the types of datasets related to the population,this paper also established a publicized multi-type scene dataset Xinzheng Airport dataset.This article uses four mainstream public data sets for verification.The experimental results show that the performance of the Multi-Aware CNN network algorithm is about 20 percentage points higher than the current best method.