基于COB-KDE融合的行人检测方法
|
摘要
核密度无需估计参数及背景特征分布,可以处理复杂背景下运动目标检测,但核函数带宽的选择一直困扰着算法的应用。针对这一问题,提出人体特征和核密度估计融合方法(characteristics of body-kernel density estimation,COBKDE),用于行人检测。通过人体特征信息选择核函数带宽,基于核密度估计提取前景,融合人体特征检测视频行人。实验结果表明,引入人体特征信息后,该方法相比于传统方法减少了核密度估计的运算量,在受到光线变化和噪声等干扰时,能准确检测行人和非行人。该方法可以推广到车辆、动物的检测。
Kernel density can deal with moving object detection in complex background without estimating parameters and background feature distribution.However,the consistency of the bandwidth of the kernel function troubles the application of the algorithm.To solve this problem,The COB-KDE(characteristics of body-kernel density estimation)method was proposed.The bandwidth of kernel function was selected using human body feature information and the foreground information based on kernel density estimation was extracted.Human body feature was fused to detect video pedestrians.Experimental results show that the introduction of human body feature information reduces the computational complexity of nuclear density estimation compared with traditional methods,and it can accurately detect pedestrians and non-pedestrians when light changes and noise disturbing.This method can be extended to vehicle and animal detection.
Kernel density can deal with moving object detection in complex background without estimating parameters and background feature distribution.However,the consistency of the bandwidth of the kernel function troubles the application of the algorithm.To solve this problem,The COB-KDE(characteristics of body-kernel density estimation)method was proposed.The bandwidth of kernel function was selected using human body feature information and the foreground information based on kernel density estimation was extracted.Human body feature was fused to detect video pedestrians.Experimental results show that the introduction of human body feature information reduces the computational complexity of nuclear density estimation compared with traditional methods,and it can accurately detect pedestrians and non-pedestrians when light changes and noise disturbing.This method can be extended to vehicle and animal detection.
引文
[1]Shao Ling,Liu Yan,Zhang Jianguo.Human action segmentation and recognition via motion and shape analysis[J].Pattern Recognition Letters,2012,33(4):438-445.
[2]Takamitsu Matsubara,Sang-Ho Hyon,Jun Morimoto.Realtime stylistic prediction for whole-body human motions[J].Neural Networks,2012,25:191-199.
[3]Chen Dong,Ren Shaoqing,Wei Yichen,et al.Joint cascade face detection and alignment[C]//ECCV,2014:109-122.
[4]Elgarnmal A,Harwood D,Davis L.Nonparametric model for background subtraction[C]//Proceedings of 16th European Conference on Computer Vision.Ireland:Dublin,2014:751-767.
[5]LI Yanping,LIN Jianhui,YANG Ningxue.A moving object tracking algorithm based on SIFT characteristic optical flow[J].Computer Science,2015,11(42):305-309(in Chinese).[李艳萍,林建辉,杨宁学.一种基于SIFT特征光流的运动目标跟踪算法[J].计算机科学,2015,11(42):305-309.]
[6]Wang L,Hu W,Tan T.Recent developments in human motion analysis[J].Pattern Recognition,2013,36(3):585-601.
[7]Zhang Li,Yu Haibin.Human tracking method under occusion with two cameras[J].Journal of Image and Graphics,2015,16(4):606-612.
[8]Sun Yan,Junzo Watada,Weng Peifen.Pedestrian recognition and tracking method based on particle swarm optimization[J].Computer Engineering and Design,2016,32(3):988-994.
[9]LI Zhi,XIE Qiang.Moving object tracking based on improved particle filter[J].Computer Science,2014,41(2):232-235(in Chinese).[李志,谢强.一种基于改进粒子滤波的运动目标跟踪[J].计算机科学,2014,41(2):232-235.]
[10]Yu Jing,Xu Dongbin,Liao Qingmin.Image defogging:A survey[J].Journal of Image and Graphics,2015,16(9):1561-1576.
[11]Jing Ting,Zhou Fugen,Bai Xiangzhi.Moving object detection in airborne video using kernel density estimation[J].Infrared and Laser Engineering,2011,40(1):153-157.
[12]Fan Zhihui.Research on motion detection and object lacking based on spatio-temporal relationship learning[D].Hefei:Xidian University,2016.
[13]Ouyang W,Wang X.Joint deep learning for pedestrian detection[C]//IEEE International Conference on Computer Vision.Sydney,2013:2056-2063.
[2]Takamitsu Matsubara,Sang-Ho Hyon,Jun Morimoto.Realtime stylistic prediction for whole-body human motions[J].Neural Networks,2012,25:191-199.
[3]Chen Dong,Ren Shaoqing,Wei Yichen,et al.Joint cascade face detection and alignment[C]//ECCV,2014:109-122.
[4]Elgarnmal A,Harwood D,Davis L.Nonparametric model for background subtraction[C]//Proceedings of 16th European Conference on Computer Vision.Ireland:Dublin,2014:751-767.
[5]LI Yanping,LIN Jianhui,YANG Ningxue.A moving object tracking algorithm based on SIFT characteristic optical flow[J].Computer Science,2015,11(42):305-309(in Chinese).[李艳萍,林建辉,杨宁学.一种基于SIFT特征光流的运动目标跟踪算法[J].计算机科学,2015,11(42):305-309.]
[6]Wang L,Hu W,Tan T.Recent developments in human motion analysis[J].Pattern Recognition,2013,36(3):585-601.
[7]Zhang Li,Yu Haibin.Human tracking method under occusion with two cameras[J].Journal of Image and Graphics,2015,16(4):606-612.
[8]Sun Yan,Junzo Watada,Weng Peifen.Pedestrian recognition and tracking method based on particle swarm optimization[J].Computer Engineering and Design,2016,32(3):988-994.
[9]LI Zhi,XIE Qiang.Moving object tracking based on improved particle filter[J].Computer Science,2014,41(2):232-235(in Chinese).[李志,谢强.一种基于改进粒子滤波的运动目标跟踪[J].计算机科学,2014,41(2):232-235.]
[10]Yu Jing,Xu Dongbin,Liao Qingmin.Image defogging:A survey[J].Journal of Image and Graphics,2015,16(9):1561-1576.
[11]Jing Ting,Zhou Fugen,Bai Xiangzhi.Moving object detection in airborne video using kernel density estimation[J].Infrared and Laser Engineering,2011,40(1):153-157.
[12]Fan Zhihui.Research on motion detection and object lacking based on spatio-temporal relationship learning[D].Hefei:Xidian University,2016.
[13]Ouyang W,Wang X.Joint deep learning for pedestrian detection[C]//IEEE International Conference on Computer Vision.Sydney,2013:2056-2063.