| Crowd density estimation has important application in many fields,such as public safety management,public space design,and public transportation scheduling management.With the widespread deployment of a large number of surveillance systems,more and more attention has been paid to the estimation of crowd density distribution with multiple surveillance videos based on computer vision methods,and the research has made certain progress.However,existing state-of-the-art single-view crowd density estimation models are difficult to estimate crowd density efficiently because of the large number of parameters of networks.In practical application,it needs a trade-off between the accuracy and the efficiency of the model for deployment.And existing crowd density estimation methods use the homography transformation or the geometric model of camera imaging for geospatial mapping when collaborating with multiple videos.It relies on high-precision camera parameters.But in some monitoring scenes,it is difficult to obtain the highprecision camera parameters,which restricts the application of the existing methods in many scenes.Aiming to solve the limitations above,the main research contents of this thesis are as follows:(1)This thesis optimizes a lightweight CSRNet model based on the Inception module.To obtain acceptable crowd density estimation accuracy with fewer parameters,this work reduces the number of model parameters significantly by designing a multiscale feature extraction module based on the Inception module to replace several layers of CSRNet and removing CNN layers which channel number is too large.The test result on the public UCF-QNRF dataset shows that compared with the existing lightweight crowd density estimation model with the same level of parameters,the crowd counting accuracy is improved by our lightweight CSRNet model with multi-scale feature extraction module,which reduces by 13.2% on the average absolute error and reduces by 10.8% on the root mean square error.(2)This thesis proposes a method for estimating the geographical distribution of wide-area crowd density based on georeferencing.To solve the problem that the existing crowd density distribution estimation methods joint with multiple videos are difficult to estimate the geographical distribution of crowd density without camera parameters and that with limited control points,this study uses the single-view crowd density estimation model to estimate the crowd density distribution of multiple video frames,then avoids the geometric process of camera imaging,uses the georeferencing method to map the crowd density distribution to geo-space with limited control points,and the geographical distribution of the wide-area crowd density is estimated after spatial fusion.Furthermore,this study proposes a wide-area crowd density estimation dataset that contains dense crowds to verify the proposed method. |
