第 10 期, 等 : RG-D ), ).,, [-3] [4-5]. [6],,, [7]., Gordon [8],, RG-D, Kinect Xtion, RG-D (),,, RG-D.,,. Campani [9],, Komogorov [10] () [11-1],

Size: px

Start display at page:

Download "第 10 期, 等 : RG-D 1817. 1), ).,, [-3] [4-5]. [6],,, [7]., Gordon [8],, RG-D, Kinect Xtion, RG-D (),,, RG-D.,,. Campani [9],, Komogorov [10] () [11-1],"

乃询从
7 years ago
Views:

1 第 7 卷第 10 期计算机辅助设计与图形学学报 Vo. 7 No 年 10 月 Journa of Computer-Aided Design & Computer Graphics Oct. 015 基于时空测地线传播的 RG-D 视频分割王斌 1), 陈文拯 1), 钟凡 1)*, 屠长河 1), 秦学英 1) ), 彭群生 1) ( 山东大学计算机科学与技术学院济南 50101) ) ( 浙江大学 CAD&CG 国家重点实验室杭州 31007) (zhongfan@sdu.edu.cn) : 针对前景和背景深度交叠或相机运动时基于深度统计的传统视频分割算法中存在的问题, 提出一种基于时空测地线的方法, 并证明该方法适合基于深度的视频分割. 首先使用基于运动检测的方式进行初始化 ; 然后使用基于特征点选择方式定义种子结点, 特征点匹配方式构建时域链接, 空间上 8 邻域像素连接形成空域链接, 在连续两帧之间构建时空测地线传播图 ; 最后在时空测地线传播图上使用泛化测地线距离变换将前一帧的分割结果传播到当前帧, 并自适应地在传播和检测间切换消除累计误差. 实验结果表明, 该方法能够在复杂场景和相机运动情形下输出稳定的分割结果. : 视频分割 ; 深度图 ; 测地距离 ; 分割传播 :TP RG-D Video Segmentation via Geodesic Spatio-Tempora Propagation Wang in 1), Chen Wenzheng 1), Zhong an 1)*, Tu Changhe 1), Qin Xueying 1), and Peng Qunsheng ) 1) (Schoo of Computer Science and Technoogy, Shandong University, Ji nan 50101) ) (State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou 31007) Abstract: Traditiona video segmentation methods based on depth statistics often fai in case of moving camera or overap between depth ranges of foreground and background. In this paper we propose a geodesic-based method that is suitabe for RG-D video segmentation. The segmentation process is initiaized by motion detection, then for each two consecutive frames, a geodesic spatio-tempora graph is constructed, with seed nodes seected based on image feature detection, tempora inks buit via feature matching and the 8 spatia neighborhoods of pixes used as spatia inks. The segmentation resut of previous frame then is propagated to the current frame via geodesic spatio-tempora propagation, which is conducted efficienty by generaized geodesic distance transform. Accumuated errors are eiminated by aternativey aunching the process of propagation and motion detection. Experiments demonstrate the robust segmentation resuts for videos of compex scenes and moving camera. Key words: video segmentation; depth map; geodesic distance; segmentation propagation, [1] 基金项目 ( , , , ); (S013DX011). 王斌 (1990),,, ; 陈文拯 (199),,, ; 钟凡 (198),,,, ; 屠长河 (1968),,,,, ; 秦学英 (1966),,,,, ; 彭群生 (1947),,,,, CC.

cn) : 针对前景和背景深度交叠或相机运动时基于深度统计的传统视频分割算法中存在的问题, 提出一种基于时空测地线的方法, 并证明该方法适合基于深度的视频分割.

2 第 10 期, 等 : RG-D ), ).,, [-3] [4-5]. [6],,, [7]., Gordon [8],, RG-D, Kinect Xtion, RG-D (),,, RG-D.,,. Campani [9],, Komogorov [10] () [11-1], TO. Crabb [11], Wang [1],,,.,,.. (), (),, 1 时空测地线传播,, [7, 13].,,,,, :, 1.1 时空传播图,, 1., [14]., 1,, :. 1.,,.,,, Shi [15] x 1. min, Shi [15] min, 1010., min

3 1818 计算机辅助设计与图形学学报第 7 卷...,,,., K, K 5. s, (Sum of Squared Differences, SSD), ( s, x) SSD ( s, x) (1 )SSD ( s, x). c c d d, x ; s 11 ; SSDc SSDd, SSD 55; c,, c 0.6. (, s x) (, s x),. K, K.,., d min, d min 5.,, K,., t ( s,x) [ ( s,x)] p ;, ; p p,, [3,10], p [0.5,1],., 8..,,,,,,,,,., s ( x, y) xy Dx Dy (1 xy ) I x I y., D I ;, xy,, xy Cx Cy r min(, ) 1 e ; C, [ x ] 1 Cx [ x] Dx Dx (1) ; [] {0,1}. Kinect, 0. (1), e,, WW( W a. b. c. d. e.

4 第 10 期, 等 : RG-D ) D, c, D W. D, d. 00, D D, x, C x [ x ]. 1. 泛化测地线距离变换 [14,16] [17] 3,. 1,. ij, i j, ij. E, E x. Gx ( ;, E) min( dxx (, ) Ex ( )) ;, x d( x,x) min ij, [ x,x] [ x,x] ( i, j ) [ x,x ] x x, ( i, j) [18]. 1.1, ( s ( i, j) ), E, t E ( x) min{{ ( s, x)},in}, {,}; ss, S, IN,. 1.1, x,., E ( x) x ;, E ( x) IN. Gx ( ;, E ) Gx ( ;, E ),. Gx ( ;, E ) Gx ( ;, E ), x ;,., Kinect,,,, ij ij,. 基于时空测地线传播的视频分割,,,.1 基于运动检测的自动初始化...,,, Ŝ, t t 1 t t1, ˆ x x x x x M I I D D x S x. T 0, Ê ( T0 M x ) 0 ˆ E e, ˆ x x S and Mx T0 ; IN, otherwise Ê ( Mx T1 ) 1 ˆ E e, ˆ x x S and M T. IN, otherwise, T1, T0 300, T1 1000, 1000, ˆ 1., (1), (1) C ˆ [ x ], x.,, Ê, Ê ˆ, 1.,, x 1

, E ( x) x ;, E ( x) IN. Gx ( ;, E ) Gx ( ;, E ),. Gx ( ;, E ) Gx ( ;, E ), x ;,., Kinect,,,, ij ij,. 基于时空测地线传播的视频分割,,,.1 基于运动检测的自动初始化.

5 180 计算机辅助设计与图形学学报第 7 卷, 1, { x SM ˆ, x T1 }, Sˆ, ;., 0.8max 0.9,, max ( 5 s).,,. 传播与检测切换策略,.,.,,.1 ;,., max,.,, 3 实验及结果分析 Kinect, Kinect, GHz CPU, 4 G. 3 4, 1,, 4, [11].,,, a. b. c. 4 5, () (), [1], [1], 5.,,,,,, a. b. c. d 结果和对比, a. b. c. 5

,, 3 实验及结果分析 Kinect, Kinect, 640480..9 GHz CPU, 4 G.

6 第 10 期, 等 : RG-D ,.,.. 8.,, 9. Kinect,,, 计算性能 1 /s, 表 1 测试视频相关数据和视频分割速度统计 /s /(/s) 分析,.,,,,. 10a, ( 1 )( ) ; 10b, 10c; 10c 算法局限性,,,, 8 a. b. c. 10

7 18 计算机辅助设计与图形学学报第 7 卷,,, Kinect,,,, TO,, [19-0]. 4 结论 RG-D..,,,.,, RG-D., Kinect, RG-D 参考文献 (References): [1] Zhang Yijiang, Qin Xueying, Juien Pettré, et a. Inserting virtua characters into ive video of rea scenes[j]. Journa of Computer-Aided Design & Computer Graphics, 011, 3(1): (in Chinese) (,, Juien Pettré,. [J]., 011, 3(1): ) [] Han, Davis L S. Density-based mutifeature background subtraction with support vector machine[j]. IEEE Transactions on Pattern Anaysis and Machine Inteigence, 01, 34(5): [3] Cheng L, Gong M L, Schuurmansc D, et a. Rea-time discriminative background subtraction[j]. IEEE Transactions on Image Processing, 011, 0(5): [4] Criminisi A, Cross G, ake A, et a. iayer segmentation of ive video[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Los Aamitos: IEEE Computer Society Press, 006: [5] Yin P, Criminisi A, Winn J, et a. iayer segmentation of webcam videos using tree-based cassifiers[j]. IEEE Transactions on Pattern Anaysis and Machine Inteigence, 011, 33(1): 30-4 [6] Zhong, Qin X Y, Peng Q S. Transductive segmentation of ive video with non-stationary background[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Los Aamitos: IEEE Computer Society Press, 010: [7] Zhong, Qin X Y, Peng Q S, et a. Discontinuity-aware video object cutout[j]. ACM Transactions on Graphics, 01, 31(6): Artice No.175 [8] Gordon G, Darre T, Harvie M, et a. ackground estimation and remova based on range and coor[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Los Aamitos: IEEE Computer Society Press, 1999, : [9] Campani M, Sagado L. ackground foreground segmentation with RG-D Kinect data: an efficient combination of cassifiers[j]. Journa of Visua Communication and Image Representation, 014, 5(1): [10] Komogorov V, Criminisi A, ake A, et a. i-ayer segmentation of binocuar stereo video[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Los Aamitos: IEEE Computer Society Press, 005, : [11] Crabb R, Tracey C, Puranik A, et a. Rea-time foreground segmentation via range and coor imaging[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshops. Los Aamitos: IEEE Computer Society Press, 008: 1-5 [1] Wang L, Gong M L, Zhang C X, et a. Automatic rea-time video matting using time-of-fight camera and mutichanne Poisson equations[j]. Internationa Journa of Computer Vision, 01, 97(1): [13] ai X, Wang J, Simons D, et a. Video SnapCut: robust video object cutout using ocaized cassifiers[j]. ACM Transactions on Graphics, 009, 8(3): Artice No. 70 [14] ai X, Sapiro G. A geodesic framework for fast interactive image and video segmentation and matting[c] //Proceedings of the 11th IEEE Internationa Conference on Computer Vision. Los Aamitos: IEEE Computer Society Press, 007: 1-8 [15] Shi J, Tomasi C. Good features to track[c] //Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Los Aamitos: IEEE Computer Society Press, 1994: [16] Criminisi A, Sharp T, ake A. GeoS: geodesic image segmentation[c] //Proceedings of the 10th European Conference on Computer Vision. Heideberg: Springer, 008: [17] Criminisi A, Sharp T, Rother C, et a. Geodesic image and video editing[j]. ACM Transactions on Graphics, 010, 9(5): Artice No. 134 [18] Toivanen P J. New geodosic distance transforms for gray-scae images[j]. Pattern Recognition Letters, 1996, 17(5): [19] Park J, Kim H, Tai Y W, et a. High quaity depth map upsamping for 3D-TO cameras[c] //Proceedings of IEEE Internationa Conference on Computer Vision. Los Aamitos: IEEE Computer Society Press, 011: [0] Min D, Lu J, Do M N. Depth video enhancement based on weighted mode fitering[j]. IEEE Transactions on Image Processing, 01, 1(3):

, 011, 3(1): 185-191) [] Han, Davis L S. Density-based mutifeature background subtraction with support vector machine[j].

何秋琳张立春视觉学习研究进展视觉注意视觉感知

何秋琳张立春视觉学习研究进展视觉注意视觉感知第卷第期年月开放教育研究何秋琳张立春华南师范大学未来教育研究中心广东广州随着图像化技术和电子媒体的发展视觉学习也逐步发展为学习科学的一个研究分支得到研究人员和教育工作者的广泛关注基于此作者试图对视觉学习