Dense Crowd Localization Method Based on Feature Denoising
The difficulty of dense crowd location analysis is to obtain accurate prediction of individual targets in the image.In dense crowd scenes,small target objects and other objects in the foreground image have characteristic noises caused by mutual occlusion and interference.The human head features learned by traditional crowd location methods are easily affected by feature noise,which may lead to weak discrimination of human head features and inaccurate boundary information acquisition.To solve the above prob-lems,a class aware feature denoising method is proposed for crowd location,which uses the idea of semantic feature decoupling to suppress feature noise and thus to enhance the detection of independent heads.Different from the traditional pixel domain denoising methods aiming at improving image visual quality,the proposed method will denoise multi-scale features in the feature space,pro-mote the model to learn more about target features,and suppress interference features.Through the semantic decoupling of fore-ground target features and background features,the response of head features and background features is enhanced and weakened,re-spectively and the detection performance of independent individual targets can be improved.Experimental results show that the aver-age F1 values of the proposed method on Shanghai Tech,UCF-QNRF and NWPU crowd dense population data sets are 81.2％,72.4％and 77.1％respectively,which shows that it improves the performance of dense population localization.
国家科技期刊平台
NSTL
万方数据
