基于背景差与HOG-LBP的快速运动人体检测  

Rapid Motion Human Detection Based on Background Subtraction and HOG- LBP

作　　者：张慧[1] 郑爱华[1] 涂铮铮[1] 罗斌[1] 

机构地区：[1]安徽大学计算机科学与技术学院,安徽合肥230601

出　　处：《计算机技术与发展》2015年第2期47-51,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61211130309);安徽高校省级自然科学研究重大项目(KJ2010ZD10)

摘　　要：针对单目静止摄像机近距离监控的情形,结合运动目标外接矩形长宽比,提出一种HOG特征联合LBP特征并通过PCA降维的快速运动人体检测算法。该方法包含两个步骤:运动目标提取和运动人体检测。使用帧差与背景差相结合的方法提取运动目标,帧差用于更新背景,背景差用于提取运动目标。运动目标判别即人体检测分为两个部分:单运动人体检测以及多运动人体检测。首先根据运动目标外接矩形的长宽比,把目标分为单目标以及多目标;然后,根据肤色的分布判断单个行人。对于多目标,提取HOG-LBP特征,用PCA降维,结合线性SVM进行群人目标判定。实验结果表明,该方法不仅提高了人体检测速度,还提高了人体检测率。For the case of close monitoring by monocular static camera, propose a rapid motion human detection algorithm combined with target movement external rectangular length-width ratio, including two steps of moving objects extraction and moving human detection. The moving target extraction is implemented by combining the frame difference and background subtraction, where the frame difference is used to update the background and background subtraction is used to extract the moving target. Moving target discrimination that is human detection can be divided into two parts, the single movement human detection and the crowd detection. Firstly, according to the empirical value of length-width ratio of bounding rectangle of the extracted moving target, the target is divided into single-object and multi-ob- ject. Then, single-object is determined according to the skin color distribution. For multi-object, HOG-LBP features are extracted, fol- lowed by PCA （Principal Component Analysis） dimensionality reduction. Then the multi-object discrimination is approached by combi- ning with the linear SVM. The experimetal results show the remarkable performance of this method is improved on both detection rate and efficiency.

关 键 词：运动人体检测 梯度方向直方图 背景差 局部二元模式 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]