基于AdaBoost和SVM的人体检测
|
摘要
随着计算机技术和人工智能技术的不断发展,目标检测已成为计算视觉领域一个热点研究课题。其中人体类检测是目标检测的一个重要分支,包括人体识别、人体跟踪、步态识别等方面,在人机交互和智能视频监控领域都有着广泛的应用前景。
人体检测作为人体类计算机视觉重要组成部分,是人体轮廓提取、人体动态分析、人体行为鉴别等相关领域研究的首要步骤。人体具有非刚性的特性,在公共场合中,由于存在人体姿态不一、衣着各异且背景复杂、光照条件多变等因素的影响,给人体检测的实现带来困难和挑战。目前主流的人体检测方法有两大分支:基于弱分类器学习法(AdaBoost)和基于支持向量机法(SVM)。其中基于弱学习的方法是通过对大量样本进行训练实验后,提取其特征,建立标准的人体统计学模型,从而分辨出候选目标中的人体;该方法理论上能将人体识别分类的性能指数趋近于最大化,但以样本数量趋近于无穷大为前提条件。而支持向量机的方法是基于结构风险最小化的原理,通过较少的样本训练得到性能较好的分类器,但存在检测耗时大的缺点。
通过对目前国内外主流人体检测方法的深入研究,并分析了各种方法中优点与不足,本文提出了一种基于改进的AdaBoost并结合SVM的算法。
1.在AdaBoost现有算法的基础上,改进其矩形特征的输出形式,使其能够更好的分辨候选目标边缘区域与平坦区域,而且特征数量大幅减少,算法计算速度得到提高;另外,构造了链式梯度特征,该特征能够根据样本边缘纹理分布情况进行自动合并生长,这样检测时特征子窗将集中于人体边缘。
2.针对弱分类器的区间划分对特征波动过于敏感的现象,提出将邻近区间模糊化划分,从而提高了算法的稳定性。
3.为避开AdaBoost算法存在的误检问题,将级联式的SVM分类器组合到算法中,通过SVM的二次检测,能够较好的剔除误检。
With the development of technology of computer and artificial intelligence , object detection becomes a pop task in the field of computer vision.Human detection, as a significant branch in the area of object detection, includes human recognition, human tracking and gait recognition. It will be widely applied in the scopes of interaction of human computer and video supervising.
As an important constitution of computer vision, human detection is the first step for these fields: human edge extraction, human dynamic analysis and human action recognition.However, human detection is not a uncomplicated problem for these reasons: different pose and dress of human body, complex background and illumination. Currently, weak learning algorithm of AdaBoost and Support Vector Machine algorithm are prevalent methods for human detection. The model of human is established after the training for a great quantity of samples, and human body from the candidate object is recognised by this model, this is the first method.It has a powerful classfication ability, but it also demands a great number of samples.Besides, the SVM method is based on the principle of structure risk minimization. Lower number of samples is required in this method and better effect for its result of detection,however, time wasting is a disadvantage of SVM algorithm.
By the research for chief method of human dection and the analysis for their advantages and disadvantages, this thesis proposes a new algorithm which based on improved AdaBoost and SVM.
1. In order to distinguish the edge area and smooth area from candidate object and reduce the quantity of features, the output form of rectangular feature is improved.Besides, chained feature is structured which can combine and self-grow along the edge of an object sample. In this way, the sub-windows of features always focus on the edge of human body.
2. A obscure method is applied in interval division. This method can solve the sensitive problem of weak classfication. By the way,the stability of AdaBoost algorithm is boosted up.
3. Cascade SVM classfication is merged with improved AdaBoost classfication which is executed after AdaBoost. As a result, the false alarms can be removed.
人体检测作为人体类计算机视觉重要组成部分,是人体轮廓提取、人体动态分析、人体行为鉴别等相关领域研究的首要步骤。人体具有非刚性的特性,在公共场合中,由于存在人体姿态不一、衣着各异且背景复杂、光照条件多变等因素的影响,给人体检测的实现带来困难和挑战。目前主流的人体检测方法有两大分支:基于弱分类器学习法(AdaBoost)和基于支持向量机法(SVM)。其中基于弱学习的方法是通过对大量样本进行训练实验后,提取其特征,建立标准的人体统计学模型,从而分辨出候选目标中的人体;该方法理论上能将人体识别分类的性能指数趋近于最大化,但以样本数量趋近于无穷大为前提条件。而支持向量机的方法是基于结构风险最小化的原理,通过较少的样本训练得到性能较好的分类器,但存在检测耗时大的缺点。
通过对目前国内外主流人体检测方法的深入研究,并分析了各种方法中优点与不足,本文提出了一种基于改进的AdaBoost并结合SVM的算法。
1.在AdaBoost现有算法的基础上,改进其矩形特征的输出形式,使其能够更好的分辨候选目标边缘区域与平坦区域,而且特征数量大幅减少,算法计算速度得到提高;另外,构造了链式梯度特征,该特征能够根据样本边缘纹理分布情况进行自动合并生长,这样检测时特征子窗将集中于人体边缘。
2.针对弱分类器的区间划分对特征波动过于敏感的现象,提出将邻近区间模糊化划分,从而提高了算法的稳定性。
3.为避开AdaBoost算法存在的误检问题,将级联式的SVM分类器组合到算法中,通过SVM的二次检测,能够较好的剔除误检。
With the development of technology of computer and artificial intelligence , object detection becomes a pop task in the field of computer vision.Human detection, as a significant branch in the area of object detection, includes human recognition, human tracking and gait recognition. It will be widely applied in the scopes of interaction of human computer and video supervising.
As an important constitution of computer vision, human detection is the first step for these fields: human edge extraction, human dynamic analysis and human action recognition.However, human detection is not a uncomplicated problem for these reasons: different pose and dress of human body, complex background and illumination. Currently, weak learning algorithm of AdaBoost and Support Vector Machine algorithm are prevalent methods for human detection. The model of human is established after the training for a great quantity of samples, and human body from the candidate object is recognised by this model, this is the first method.It has a powerful classfication ability, but it also demands a great number of samples.Besides, the SVM method is based on the principle of structure risk minimization. Lower number of samples is required in this method and better effect for its result of detection,however, time wasting is a disadvantage of SVM algorithm.
By the research for chief method of human dection and the analysis for their advantages and disadvantages, this thesis proposes a new algorithm which based on improved AdaBoost and SVM.
1. In order to distinguish the edge area and smooth area from candidate object and reduce the quantity of features, the output form of rectangular feature is improved.Besides, chained feature is structured which can combine and self-grow along the edge of an object sample. In this way, the sub-windows of features always focus on the edge of human body.
2. A obscure method is applied in interval division. This method can solve the sensitive problem of weak classfication. By the way,the stability of AdaBoost algorithm is boosted up.
3. Cascade SVM classfication is merged with improved AdaBoost classfication which is executed after AdaBoost. As a result, the false alarms can be removed.
引文
[1]Collins R et al. A system for video surveillance and monitoring:VSAM final report[C]. Technical Report on CMU-RI-TR-00-12, 2000: 102-107.
[2]Haritaoglu I, Harwood D and Davis L. W4:real-time surveillance of people and their activities[J]. IEEE Trans Pattern Analysis and Machine Intelligence, 2000, 22(8): 809-830.
[3]Pedeo F Felzenszwalb and Daniel P Huttenlocher. Effcient Matching of Pictorial Structures[C]. Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2000: 66-73.
[4]Greg Mori and Jitendra Malik. Estimating Human Body Con-gurations using Shape Context Matching[C]. Euro Conference on Computer Vision LNCS 2352, 2002: 666-680.
[5]Lipton A, Fujiyoshi H, Patil R. Moving target classfication and tracking from real-time video[C]. IEEE Workshop on Applications of Computer Vision, 1998: 8-14.
[6]Shi J and Malik J. Normalized cuts and image segmentation[J]. IEEE Trans.Pattern Analysis And Machine Intelligence, 2000, 22(8): 888-905.
[7]曹云忠.序列图像处理的快速人体检测方法[J].西华大学学报(自然科学版), 2004, 43(3): 53-57.
[8] Yunhui He. Modified Gradient Linear Discriminant Analysis for Face Recognition[C]. IEEE International Symposium on Knowledge Acquisition and Modeling Workshop, 2008, 563-566.
[9]郭志波.人脸快速检测和特征抽取方法的研究[D].南京:南京理工大学, 2007.
[10] Turk M, Pentland A. Face recognition using eigenface[C]. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1991: 586-591.
[11] Rowley H A, Baluja S, Kanade T. Neural network-based face detection [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998,20(1):23-28.
[12] El-Khamy, Abdel-Alim, Saii O A. Neural network face recognition using statistical feature extraction[C]. Radio Science Conference. 17th NRSC2000 Seventeenth National, 2000: C31/1-C31/8.
[13] Lawrence S, Giles C L, Ah Chung Tsoi. Face recognition: A Convolutional Neural-Network Approach[J]. IEEE Transactions on Neural Networks, 1997, 8(1): 98-113.
[14]杜树新,吴铁军.模式识别中的支持向量机方法[J].浙江大学学报(工学版), 2003, 3(5): 522-527.
[15]赵博,李永忠,杨鸽,徐静.改进SVM在入侵检测中的应用研究[J].计算机工程与应用,2009, 45(17): 102-104.
[16]叶苗,王勇,麦范金.基于改进的层叠SVM模型的入侵检测技术[J].微计算机信息, 2008, 24(3): 78-79.
[17]沈谦,李树涛,伍君.基于主分量分析和支持向量机的人脸检测[J].计算机与数学工程, 2004, 33(4): 56-58.
[18]赵伟.基于SVM的入侵检测研究[D].北京:北京交通大学, 2007.
[19]陈锻生,刘政凯.利用图像颜色及其边缘直方图特征的SVM人脸检测[J].小型微型计算机系统, 2005, 26(12): 2194-2199.
[20]陈蕾,黄贤武,仲兴荣,王加俊.基于支持向量机的运动目标中人脸检测[J].微电子学与计算机, 2005, 22(6): 19-22.
[21]李同治,丁晓青,王生进.利用级联SVM的人体检测方法[J].中国图象图形学报, 2008, 13(3): 566-570.
[22] Li S Z, Zhenqiu Zhang. FloatBoost learning and statistical face detection[J]. Pattern Analysis and Machine Intelligence, 2004, 26(9): 1112-1129.
[23]武勃,黄畅,艾海舟,劳世竑.基于连续AdaBoost算法的多视角人脸检测[J].计算机研究与发展, 2005, 42(9): 1612-1621.
[24]田源,于凤芹.一种改进的AdaBoost的人脸检测算法[J].江南大学学报(自然科学版), 2009, 8(3): 283-287.
[25]刘侠,李苏,李廷军.一种改进的AdaBoost算法的人脸检测分类器[J].空军工程大学学报(自然科学版), 2009, 10(2): 76-80.
[26]刘宏志.基于一种改进的AdaBoost算法的人脸检测系统的设计与实现[D].北京:北京大学, 2007.
[27]严超,王元庆.基于连续型AdaBoost算法和Cascade结构的红外人脸检测[J].光与红外, 2009, 39(11): 1246-1250.
[28]王晶,杨煜.基于边缘方向直方图的AdaBoost人脸检测[J],计算机技术与发展, 2007, 17(12): 5-7.
[29]陈春燕,章品正,罗立民.基于粒特征和连续AdaBoost的人脸检测[J].智能系统学报, 2009, 4(5): 446-452.
[30]林鹏.基于AdaBoost算法的人脸检测研究及实现[D].西安:西安理工大学, 2007.
[31]王勇,陶晓玲.分级结构的AdaBoost入侵检测方法研究[J].西安电子科技大学学报(自然科学版), 2008, 35(2): 345-350.
[32] Freund Y, Schapire R E. Experiments with a new boosting algorithm[C]. In Proceedings of the 13th International Conference on Machine Learning, 1996: 148-156.
[33] Schapire R E. The strength of weak learnability[C]. Annual Symposium on Foundations of Computer Science, 1989: 28-33.
[34] Schapire R E. A Brief Introduction to Boosting[C]. In Proceedings of the 16th International Joint Conference on Artificial Intelligence, 1999: 1401-1406.
[35] Dalal N, Triggs B. Histograms of oriented gradients for human detection[C]. Inernational Compuer Vision and Pattern Recognition, 2005: 886-893.
[36] Maolin Chen, Gengyu Ma, Seokcheol Kee. Multi-view human head detection in static images [C]. IAPR Conference on Machine Vision Application, 2005: 100-103.
[37] Viola P, Jones M. Rapid object detection using a boosted cascade of simple features[C]. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2001, 1: 511-518.
[38] Papageorgiou. C, Oren M and Poggio T A. General Framework for Object Detection[C]. International Conference on Computer Vision, 1998: 345-351.
[39] Crow F C. Summed-area tables for texture mapping[C]. International Conference on Computer Graphics and Interactive Techniques, 1984 : 207-212.
[40]邓乃杨,田英杰.数据挖掘中的新方法-支持向量机[M].北京:科学出版社, 2005, 134-147.
[41]杜喆.几类支持向量机变形算法的研究[D].西安:西安电子科技大学, 2009.
[2]Haritaoglu I, Harwood D and Davis L. W4:real-time surveillance of people and their activities[J]. IEEE Trans Pattern Analysis and Machine Intelligence, 2000, 22(8): 809-830.
[3]Pedeo F Felzenszwalb and Daniel P Huttenlocher. Effcient Matching of Pictorial Structures[C]. Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2000: 66-73.
[4]Greg Mori and Jitendra Malik. Estimating Human Body Con-gurations using Shape Context Matching[C]. Euro Conference on Computer Vision LNCS 2352, 2002: 666-680.
[5]Lipton A, Fujiyoshi H, Patil R. Moving target classfication and tracking from real-time video[C]. IEEE Workshop on Applications of Computer Vision, 1998: 8-14.
[6]Shi J and Malik J. Normalized cuts and image segmentation[J]. IEEE Trans.Pattern Analysis And Machine Intelligence, 2000, 22(8): 888-905.
[7]曹云忠.序列图像处理的快速人体检测方法[J].西华大学学报(自然科学版), 2004, 43(3): 53-57.
[8] Yunhui He. Modified Gradient Linear Discriminant Analysis for Face Recognition[C]. IEEE International Symposium on Knowledge Acquisition and Modeling Workshop, 2008, 563-566.
[9]郭志波.人脸快速检测和特征抽取方法的研究[D].南京:南京理工大学, 2007.
[10] Turk M, Pentland A. Face recognition using eigenface[C]. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1991: 586-591.
[11] Rowley H A, Baluja S, Kanade T. Neural network-based face detection [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998,20(1):23-28.
[12] El-Khamy, Abdel-Alim, Saii O A. Neural network face recognition using statistical feature extraction[C]. Radio Science Conference. 17th NRSC2000 Seventeenth National, 2000: C31/1-C31/8.
[13] Lawrence S, Giles C L, Ah Chung Tsoi. Face recognition: A Convolutional Neural-Network Approach[J]. IEEE Transactions on Neural Networks, 1997, 8(1): 98-113.
[14]杜树新,吴铁军.模式识别中的支持向量机方法[J].浙江大学学报(工学版), 2003, 3(5): 522-527.
[15]赵博,李永忠,杨鸽,徐静.改进SVM在入侵检测中的应用研究[J].计算机工程与应用,2009, 45(17): 102-104.
[16]叶苗,王勇,麦范金.基于改进的层叠SVM模型的入侵检测技术[J].微计算机信息, 2008, 24(3): 78-79.
[17]沈谦,李树涛,伍君.基于主分量分析和支持向量机的人脸检测[J].计算机与数学工程, 2004, 33(4): 56-58.
[18]赵伟.基于SVM的入侵检测研究[D].北京:北京交通大学, 2007.
[19]陈锻生,刘政凯.利用图像颜色及其边缘直方图特征的SVM人脸检测[J].小型微型计算机系统, 2005, 26(12): 2194-2199.
[20]陈蕾,黄贤武,仲兴荣,王加俊.基于支持向量机的运动目标中人脸检测[J].微电子学与计算机, 2005, 22(6): 19-22.
[21]李同治,丁晓青,王生进.利用级联SVM的人体检测方法[J].中国图象图形学报, 2008, 13(3): 566-570.
[22] Li S Z, Zhenqiu Zhang. FloatBoost learning and statistical face detection[J]. Pattern Analysis and Machine Intelligence, 2004, 26(9): 1112-1129.
[23]武勃,黄畅,艾海舟,劳世竑.基于连续AdaBoost算法的多视角人脸检测[J].计算机研究与发展, 2005, 42(9): 1612-1621.
[24]田源,于凤芹.一种改进的AdaBoost的人脸检测算法[J].江南大学学报(自然科学版), 2009, 8(3): 283-287.
[25]刘侠,李苏,李廷军.一种改进的AdaBoost算法的人脸检测分类器[J].空军工程大学学报(自然科学版), 2009, 10(2): 76-80.
[26]刘宏志.基于一种改进的AdaBoost算法的人脸检测系统的设计与实现[D].北京:北京大学, 2007.
[27]严超,王元庆.基于连续型AdaBoost算法和Cascade结构的红外人脸检测[J].光与红外, 2009, 39(11): 1246-1250.
[28]王晶,杨煜.基于边缘方向直方图的AdaBoost人脸检测[J],计算机技术与发展, 2007, 17(12): 5-7.
[29]陈春燕,章品正,罗立民.基于粒特征和连续AdaBoost的人脸检测[J].智能系统学报, 2009, 4(5): 446-452.
[30]林鹏.基于AdaBoost算法的人脸检测研究及实现[D].西安:西安理工大学, 2007.
[31]王勇,陶晓玲.分级结构的AdaBoost入侵检测方法研究[J].西安电子科技大学学报(自然科学版), 2008, 35(2): 345-350.
[32] Freund Y, Schapire R E. Experiments with a new boosting algorithm[C]. In Proceedings of the 13th International Conference on Machine Learning, 1996: 148-156.
[33] Schapire R E. The strength of weak learnability[C]. Annual Symposium on Foundations of Computer Science, 1989: 28-33.
[34] Schapire R E. A Brief Introduction to Boosting[C]. In Proceedings of the 16th International Joint Conference on Artificial Intelligence, 1999: 1401-1406.
[35] Dalal N, Triggs B. Histograms of oriented gradients for human detection[C]. Inernational Compuer Vision and Pattern Recognition, 2005: 886-893.
[36] Maolin Chen, Gengyu Ma, Seokcheol Kee. Multi-view human head detection in static images [C]. IAPR Conference on Machine Vision Application, 2005: 100-103.
[37] Viola P, Jones M. Rapid object detection using a boosted cascade of simple features[C]. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2001, 1: 511-518.
[38] Papageorgiou. C, Oren M and Poggio T A. General Framework for Object Detection[C]. International Conference on Computer Vision, 1998: 345-351.
[39] Crow F C. Summed-area tables for texture mapping[C]. International Conference on Computer Graphics and Interactive Techniques, 1984 : 207-212.
[40]邓乃杨,田英杰.数据挖掘中的新方法-支持向量机[M].北京:科学出版社, 2005, 134-147.
[41]杜喆.几类支持向量机变形算法的研究[D].西安:西安电子科技大学, 2009.