基于窗口边缘梯度势能的人体遮挡多尺度检测算法  

Multi-scale Human Detection Based on Window Gradient Potential Energy with Partial Occlusion Handling

作　　者：欧阳毅[1] 张三元[1] 张引[1] 

机构地区：[1]浙江大学计算机科学与技术学院,杭州310027

出　　处：《电子与信息学报》2012年第4期858-864,共7页Journal of Electronics & Information Technology

基　　金：国家973计划项目(2009CB320804);浙江省科技厅项目(2011C24008);广东省教育部产学研结合项目(2010B090400193)资助课题

摘　　要：为提高对于图像中人体检测的准确率,该文提出窗口边缘梯度势能(Window Gradient Potential Energy,WGPE)的概念和一种基于窗口边缘梯度势能快速人体检测方法。采用稀疏-稠密梯度势能窗口集对人体进行多尺度检测过滤,缩短了检测时间。采用改进后的加权级联支持向量机训练正负样本,将遮挡情况下的人体正样本进行加权划分,以检测遮挡环境下的人体。在对检测窗口进行过滤时,该算法并不需要增加过多的计算开销。在背景较为平滑的图像中,与多尺度面向梯度直方图(HOG)和HOG-LBP(Histograms of Oriented Gradients and Local Binary Pattern)方法相比在相同的准确率下,具有较少的检测时间。实验表明在人体检测的准确率和效率方面有所提高,对于处于半遮挡情况下人体检测,准确率也有明显提高。In order to improve accuracy of the human detection,this paper proposes the conception of the Window edge of the Gradient of Potential Energy（WGPE） and a fast human detection method based on potential energy. By using sparse-dense gradient potential windows set,the detection time of the multi-scale detection can be shortened.Cascading Support Vector Machine（SVM） training using weighted positive and negative samples,the occlusion sample of the human body is weighted to detect the human body under occlusion.Filter positive in the detection window,the algorithm does not require too much computational overhead increases when the detection window is filtered.In the smooth background image,the proposed method compared to the multi-level Histograms of Oriented Gradients（HOG） detection and HOG-LBP（Local Binary Pattern） methods accuracy at the same rate, spents less testing time.Experiments show that the human detection accuracy and efficiency has increased,the case for the human body in partial occlusion detection,the accuracy rate is improved markedly.

关 键 词：人体检测 面向梯度直方图 梯度势能 级联支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]