Research On Dense Crowd Estimation Method Based On Multi-source Information Fusion

In recent years,with the continuous growth of the city population and the enrichment of cultural life,crowd-intensive sports activities in large public places are becoming frequent.The typical feature of such activities is that the crowd density increases sharply in a short time,even exceeding the load of the place,which is easy to lead to crowding,stampede and other events endangering public safety.Dense crowd estimation technology is an important research branch in the field of computer vision,which can effectively deal with similar challenges.However,most of the existing crowd density estimation methods are based on single-view and near-ground scenes,which are difficult to solve the problems of incomplete global information and blocked local information,and can not meet the counting requirements of current large view scenes.Relying on the rapid upgrade of video monitoring hardware systems,multi-view cross-coverage monitoring image data is becoming a video system in crowded places.To comply with this development trend,this paper focuses on the research of dense crowd estimation method based on multi-source information fusion,that is,using the temporal correlation and spatial complementarity between high-altitude perspective and low-altitude perspective to improve the dense crowd estimation ability of large perspective scenes.The specific research contents include the following aspects:（1）Because of the lack of applicable high and low altitude view image datasets in this research,a new large-scale multi-view crowd counting dataset called CROWD SZ is established in this paper.It contains various images under different lighting conditions,multiple viewing angle changes,and crowd occlusion.Low altitude view images can be used for crowd counting and density estimation.Multiple information sources of low altitude view can also be used to supplement the global information of high altitude,to better show the overall crowd change trend.Therefore,the dataset is of great value and challenging to the research of crowd density estimation based on multi-source information fusion.（2）For low-altitude single-perspective monitoring scene images,the problem of large deviation of dense crowd estimation is caused by uneven head size caused by camera shooting angle and crowd occlusion,this paper proposes a low altitude view crowd counting optimization network based on attention mechanism called AMNet.Based on using VGG16 as the benchmark network,we introduced the attention mechanism module to count the crowd in the low altitude view image,focusing on the head position information in the image.The accuracy and robustness of the algorithm structure are verified in several classic crowd counting datasets,Shanghaitech,UCF_CC₅0,Beijing-BRT-dataset,and self-built dataset CROWD SZ.The experimental results show that this method effectively improves the estimation accuracy and enhances the reliability of scene counting.（3）Aiming at the widespread registration problem of overlapping areas of high and low altitude view images,this paper proposes an image fusion registration method for high and low altitude view scenes.Using the spatial complementarity between high and low altitude view images,the feature points are selected to match and calculate the homography matrix.Finally,the image overlapping area registration of high altitude view and low altitude view is realized through matrix transformation,to realize the fusion of high altitude view discrete density map and low altitude view images.（4）Aiming at the problems of incomplete low altitude view coverage and unclear texture features of high altitude view crowd in large view scene dense crowd estimation,a dense crowd estimation method based on high and low altitude view image information fusion is designed in this paper.Firstly,the crowd distribution density map of the highaltitude perspective image is discretized into four density levels,which intuitively reflects the density similarity of each region.Secondly,the density information of the high-altitude view image is corrected by the crowd count information of the low-altitude view image obtained by AMNet,to obtain a more accurate global crowd estimation value.This paper further deduces the estimated number of people from a low altitude perspective according to the proportion coefficient obtained by the fusion algorithm.Comparing with the actual number of people shows the consistency of the two groups of values,and then verifies the rationality and accuracy of the dense crowd estimation algorithm.