一种基于深度HOG和LLE的人体跟踪方法  

A HUMAN BODY TRACKING METHOD BASED ON DEEP HOG AND LLE

作　　者：刘剑[1,2] 龚志恒[1] 高恩阳[1,3] 刘亚楠[1] 

机构地区：[1]沈阳建筑大学信息与控制工程学院,辽宁沈阳110168 [2]东北大学信息科学与工程学院,辽宁沈阳110004 [3]中国科学院自动化研究所,辽宁沈阳110016

出　　处：《计算机应用与软件》2014年第4期188-192,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(61272253);辽宁省住建部项目(2010-K-9-22)

摘　　要：针对传统人体跟踪方法存在的不足,提出一种深度Histograms of Oriented Gradients(HOG)和Locally Linear Embedding(LLE)相结合的跟踪方法。首先依据图像的颜色和深度信息,结合改进的HOG表达式提取人体的特征向量;再利用流行学习LLE算法对特征向量进行降维,采用欧氏距离判别法找出每帧图像人体所在区域,并对人体区域加以标记;最后,进行人体的实时跟踪。实验结果表明:降维后的人体特征数据更有助于实现跟踪,所提出的方法应用到视频图像人体跟踪中可以有效地跟踪人体,简化了数据复杂度,明显提高人体跟踪准确率。Aiming at the deficiencies of traditional human body tracking method,we propose a tracking method which combines the deep histograms of oriented gradients（ HOG） with locally linear embedding（ LLE）. First,it extracts the eigenvector of human body according to the colour and depth information of the image and in conjunction with the improved HOG expression; Then it uses the popular learning LLE algorithm to reduce the dimensionality of eigenvectors,and use Euclidean distance discriminant to find out the regions where each image frame of human body is in,and marks these regions. Finally,it makes timely human body tracking. Experimental results show that the human body feature data with dimensionality reduced is more conducive to tracking realisation,and applying the proposed method to tracking the human body in video image will be more effective,which simplifies the data complexity and significantly improves the accuracy of human body tracking.

关 键 词：HOG 人体跟踪 深度信息 LLE 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]