Crowd Counting In Complex Public Areas Based On Object Detection And Density Map

With the rapid growth of population size and social economy,people's social activities in public places are becoming more and more frequent,such as cultural entertainment,sports events,political rallies,etc.In the meantime the attendant mass incidents are also increasing,and the safety of public places becomes particularly important.Therefore,the establishment of a smart population density monitoring system,real-time monitoring and analysis of crowded people dynamically,found the crowd abnormalities,and issued warning information,which make the administrator to take appropriate measures in advance to avoid the occurrence of malignant events.It has important practical significance.The population density distribution estimate is one of the key factors in analyzing changes in the population situation.The situation in the public area is complex,and the distribution of population density has strong randomness,making it difficult for the population density estimation and situation analysis in a specific area.This paper is based on the cluster aggregation calculation of complex public areas.In order to further analyze crowd behavior,improve the accuracy of crowd density estimation in public areas,and expand the scene applicability of crowd density estimation algorithms,This paper proposes a method for crowd density estimation of complex scenes in combination with target detection and density distribution methods:According to the analysis of our application scenarios,the monitoring coverage is relatively large,the depth of the captured image is far away,and the close-range and distant-view crowds have large differences in density and are randomly distributed.For example,although the crowd density in the close-range area is relatively low and there is less crowding,the distribution is scattered;there are a large number of crowded and obstructed cases in the foreground area.This paper proposes a new EDOF(extended depth of field)scenario for this type of situation.Based on the analysis above,this paper uses a method of image segmentation guided by image depth information to divide the image into two parts(close-range spots and distant-range spots)as input for two different neural networks.Then we analyze the general features of the near-view region and distant-view region in the scene,and use a combination of target detection and density distribution to estimate the density: For the near-view region,we propose a method based on target recognition to calculate the number of low-density people.Our algorithm can count populations in real-time scenes for non-intensive scenes.It can not only output the number of people directly but also output the specific location of the individual's image.For the distant-view region,the crowd is more intensive.First,according to the neural network,the density map in the perspective is obtained.Then calculated the number of high-density scenes by integrating the density maps.The sum of the two is finally calculated by the system to obtain a global population estimate.The experimental results show that the proposed algorithm based on the combination of detection and density distribution for crowd density estimation in complex scenes has achieved satisfactory results on different data sets and can accurately estimate population density in different scenarios.It can provide real and effective data support for public safety management.