基于混合专家网络的人群定位方法OA

In recent years,stampede incidents caused by crowd gatherings have frequently occurred,leading to increased atten-tion on monitoring crowd activities in public places.As a result,crowd analysis research has gained significant social importance.The task of crowd localization,as one of the key issues in the field of crowd analysis,aims to locate the positions of individual heads within an image.However,a single crowd image often contains visual features of varying scales,such as different head sizes and varying crowd densities and so on,making it challenging for existing methods to handle these complex situations.To ad-dress the aforementioned issues,this paper leverages the strong contextual feature association capability of the Transformer to propose a novel Transformer-based crowd localization model.Additionally,a mixed expert module is introduced to assign local-ization tasks of different scales to different experts for decoupling,thereby improving the accuracy and robustness of crowd local-ization.Finally,the proposed method is compared with existing methods on three public datasets,verifying its superior perfor-mance.