高晶 陳曉臻
摘 要: 為了提高人物動(dòng)作的三維虛擬重構(gòu)和識(shí)別能力,需進(jìn)行人物動(dòng)作的特征提取和捕捉,故提出基于AR動(dòng)態(tài)圖像的人物動(dòng)作捕捉技術(shù)。采用三維動(dòng)態(tài)跟蹤識(shí)別方法進(jìn)行人物動(dòng)作的動(dòng)態(tài)圖像特征提取,利用AR虛擬現(xiàn)實(shí)技術(shù)進(jìn)行人物動(dòng)態(tài)圖像的包絡(luò)輪廓分割,繪制反應(yīng)人物動(dòng)作特征的灰度直方圖,結(jié)合Harris角點(diǎn)跟蹤檢測(cè)方法進(jìn)行人物動(dòng)作捕捉,實(shí)現(xiàn)人物動(dòng)態(tài)圖像三維虛擬成像重構(gòu)和動(dòng)作的實(shí)時(shí)捕捉。仿真結(jié)果表明,采用該方法進(jìn)行人物動(dòng)作捕捉的動(dòng)態(tài)特征匹配能力較好,對(duì)動(dòng)態(tài)特征點(diǎn)的檢測(cè)性能較高,具有較好的人物動(dòng)態(tài)圖像虛擬重構(gòu)和識(shí)別能力。
關(guān)鍵詞: AR動(dòng)態(tài)圖像; 人物動(dòng)作; 輪廓分割; 特征提??; 識(shí)別能力; 檢測(cè)方法
中圖分類號(hào): TN911.73?34; TP391 文獻(xiàn)標(biāo)識(shí)碼: A 文章編號(hào): 1004?373X(2018)08?0144?03
Abstract: To improve the 3D virtual reconstruction and recognition capability of character actions, feature extraction and capture of character actions need to be performed. Therefore, a character action capture technology based on AR dynamic images is proposed. The 3D dynamic tracking recognition method is adopted to extract dynamic image features of character actions. The AR virtual reality technology is used to segment the envelope contour of characters′ dynamic images and draw the gray histogram which reflects action features of characters. The Harris corner tracking detection method is combined to capture character actions and realize 3D virtual image reconstruction and real?time action capture of characters′ dynamic images. The simulation results show that the method for capturing character actions has good dynamic feature matching capability, good dynamic feature point detection performance, and good virtual reconstruction and recognition capability of characters′ dynamic images.
Keywords: AR dynamic image; character action; contour segmentation; feature extraction; recognition capability; detection method
隨著虛擬現(xiàn)實(shí)(Virtual Reality,VR)技術(shù)的發(fā)展,計(jì)算機(jī)圖像處理基礎(chǔ)被廣泛應(yīng)用于目標(biāo)識(shí)別、特征分析和動(dòng)態(tài)圖像重構(gòu)等領(lǐng)域。其采用虛擬現(xiàn)實(shí)成像技術(shù)進(jìn)行人物的動(dòng)態(tài)圖像重建和動(dòng)作識(shí)別,結(jié)合人體的動(dòng)態(tài)特征圖進(jìn)行特征分析,實(shí)現(xiàn)對(duì)運(yùn)動(dòng)人物動(dòng)作的三維模擬跟蹤;通過對(duì)人物動(dòng)作的實(shí)時(shí)捕捉,結(jié)合模式識(shí)別和智能特征提取技術(shù),進(jìn)行人物動(dòng)作識(shí)別;通過動(dòng)作捕捉和識(shí)別結(jié)果進(jìn)行人體的動(dòng)態(tài)特性分析[1]。相關(guān)的人物動(dòng)態(tài)特征分析技術(shù)在體育訓(xùn)練、視頻動(dòng)態(tài)監(jiān)測(cè)以及刑事偵查等領(lǐng)域具有很好的應(yīng)用價(jià)值。
傳統(tǒng)方法中,對(duì)運(yùn)動(dòng)人物動(dòng)作捕捉和圖像識(shí)別技術(shù)主要采用三維動(dòng)態(tài)流形向量跟蹤識(shí)別方法提取運(yùn)動(dòng)人體的邊緣輪廓特征,結(jié)合自適應(yīng)角點(diǎn)檢測(cè)和圖像分割技術(shù)進(jìn)行人體動(dòng)作識(shí)別。主要的動(dòng)作捕捉方法有于LLE檢測(cè)方法和支持向量機(jī)算法(SVM)等[2]。文獻(xiàn)[3]提出基于直方圖分布特征模糊學(xué)習(xí)的人物動(dòng)作三維動(dòng)態(tài)跟蹤識(shí)別方法,結(jié)合三維動(dòng)作流形特征進(jìn)行人物圖像的動(dòng)態(tài)特征分析,實(shí)現(xiàn)人物動(dòng)作捕捉,但該方法存在計(jì)算開銷大,對(duì)人物動(dòng)作捕捉的實(shí)時(shí)性不好的問題。針對(duì)上述問題,本文提出基于AR動(dòng)態(tài)圖像的人物動(dòng)作捕捉技術(shù)。
采用三維動(dòng)態(tài)跟蹤識(shí)別方法進(jìn)行人物動(dòng)作的動(dòng)態(tài)圖像特征提取,利用AR虛擬現(xiàn)實(shí)技術(shù)進(jìn)行人物動(dòng)態(tài)圖像的包絡(luò)輪廓分割,實(shí)現(xiàn)人物動(dòng)作捕捉優(yōu)化,最后進(jìn)行仿真實(shí)驗(yàn),展示了本文方法在提高人物動(dòng)作捕捉能力方面的優(yōu)越性能。
1.1 人物動(dòng)作圖像的三維動(dòng)作流形分析
為了實(shí)現(xiàn)對(duì)人物動(dòng)作實(shí)時(shí)捕捉和動(dòng)態(tài)圖像重構(gòu),首先采用實(shí)時(shí)圖像采集方法進(jìn)行人物動(dòng)態(tài)圖像采集,提取動(dòng)態(tài)人物圖像的三維動(dòng)作流形矢量,對(duì)訓(xùn)練動(dòng)作特征點(diǎn)進(jìn)行自適應(yīng)匹配,具體的流程圖如圖1所示。
結(jié)合三維動(dòng)作特征點(diǎn)的流形分割方法,得到人物動(dòng)作重構(gòu)的動(dòng)態(tài)特征量[f(z)],在人物動(dòng)態(tài)成像的成像區(qū)域進(jìn)行三維動(dòng)作流形分解,得到人物動(dòng)作的實(shí)時(shí)捕捉先驗(yàn)信息。由此得到人物動(dòng)作重構(gòu)的動(dòng)態(tài)特征量為:
1.2 三維動(dòng)態(tài)跟蹤識(shí)別
采用三維動(dòng)態(tài)跟蹤識(shí)別方法進(jìn)行人物動(dòng)作的動(dòng)態(tài)圖像特征提取,對(duì)采集的人物動(dòng)態(tài)圖像進(jìn)行動(dòng)作特征點(diǎn)定位,在邊緣幀不變的約束條件下,采用AR虛擬現(xiàn)實(shí)技術(shù)進(jìn)行人物動(dòng)態(tài)圖像的包絡(luò)輪廓分割[4?5]。在人體動(dòng)態(tài)變化下,得到動(dòng)作向量的集合描述。在仿射不變區(qū)域內(nèi),根據(jù)人體的動(dòng)作特征進(jìn)行動(dòng)態(tài)圖像的AR重構(gòu),得到灰度像素范圍區(qū)域中的人體動(dòng)作特征捕捉的銜接動(dòng)作向量,人物動(dòng)態(tài)跟蹤識(shí)別的動(dòng)態(tài)特征函數(shù)為:
在采用三維動(dòng)態(tài)跟蹤識(shí)別方法進(jìn)行人物動(dòng)作的動(dòng)態(tài)圖像特征提取的基礎(chǔ)上,進(jìn)行人物動(dòng)作捕捉算法優(yōu)化設(shè)計(jì)[6?7]。本文提出基于AR動(dòng)態(tài)圖像的人物動(dòng)作捕捉技術(shù),對(duì)人物動(dòng)作三維動(dòng)態(tài)圖像分割和塊匹配,得到人物動(dòng)作的AR基分量。根據(jù)圖像像素跟蹤點(diǎn)進(jìn)行人物動(dòng)態(tài)圖像的角點(diǎn)檢測(cè)[8],得到人物AR動(dòng)態(tài)圖像的三維動(dòng)作捕捉分析圖如圖2所示。
AR動(dòng)態(tài)圖像的人物動(dòng)作捕捉仿真實(shí)驗(yàn)建立在Matlab 7仿真軟件上,采用L3G4200D動(dòng)態(tài)成像儀進(jìn)行人物動(dòng)態(tài)圖像采集。圖像像素點(diǎn)采樣的跟蹤偏移量設(shè)定為0.024 mm,徑向偏差設(shè)定為0.125,人物動(dòng)態(tài)圖像的分塊大小為[256×256×224,]動(dòng)態(tài)特征點(diǎn)的相關(guān)系數(shù)為0.45,灰度像素值為124,歸一化相關(guān)系數(shù)值[ω=]0.23,人體動(dòng)態(tài)特征點(diǎn)定位的標(biāo)準(zhǔn)誤差系數(shù)為[a]=0.58。根據(jù)上述仿真參量設(shè)定,進(jìn)行人體動(dòng)態(tài)特征點(diǎn)檢測(cè),得到檢測(cè)結(jié)果如圖3所示。
分析圖4得知,本文方法進(jìn)行人物動(dòng)作捕捉的動(dòng)態(tài)特征匹配能力較好。為了對(duì)比算法性能,采用本文方法和傳統(tǒng)方法測(cè)試人物動(dòng)作捕捉的準(zhǔn)確性對(duì)比結(jié)果,如圖5所示。分析圖5得知,采用本文方法進(jìn)行人物動(dòng)作捕捉的準(zhǔn)確性較高,對(duì)動(dòng)態(tài)特征點(diǎn)的檢測(cè)性能較高。
本文提出基于AR動(dòng)態(tài)圖像的人物動(dòng)作捕捉技術(shù),其采用三維動(dòng)態(tài)跟蹤識(shí)別方法進(jìn)行人物動(dòng)作的動(dòng)態(tài)圖像特征提取,利用AR虛擬重構(gòu)技術(shù)進(jìn)行人物動(dòng)態(tài)重建,結(jié)合Harris角點(diǎn)跟蹤檢測(cè)方法進(jìn)行人物動(dòng)作捕捉,實(shí)現(xiàn)人物動(dòng)態(tài)圖像的動(dòng)作實(shí)時(shí)捕捉。研究表明,本文方法在進(jìn)行人物動(dòng)態(tài)圖像動(dòng)作捕捉時(shí)準(zhǔn)確性較好,識(shí)別能力較高。
參考文獻(xiàn)
[1] 段宇,侯宇. 輪式管外攀爬機(jī)器人結(jié)構(gòu)設(shè)計(jì)與動(dòng)力特性分析[J].機(jī)械設(shè)計(jì)與制造工程,2016,45(12):17?20.
DUAN Yu, HOU Yu. The structure design and dynamic characteristics analysis of the wheeled pipe climbing robot [J]. Machinery design &; manufacture, 2016, 45(12): 17?20.
[2] 肖淑蘋,賀毅岳.一種改進(jìn)的EMD 圖像信號(hào)去噪算法[J].現(xiàn)代電子技術(shù),2016,39(16):91?93.
XIAO Shuping, HE Yiyue. An improved signal denoising algorithm of EMD image [J]. Modern electronics technique, 2016, 39(16): 91?93.
[3] LONG Tengfei, JIAO Weili, HE Guojin, et al. Automatic line segment registration using Gaussian mixture model and expectation?maximization algorithm [J]. IEEE journal of selected topics in applied earth observations and remote sensing, 2014, 7(5): 1688?1699.
[4] 卞樂,霍冠英,李慶武.基于Curvelet變換和多目標(biāo)粒子群的混合熵MRI圖像多閾值分割[J].計(jì)算機(jī)應(yīng)用,2016,36(11):3188?3195.
BIAN Le, HUO Guanying, LI Qingwu. Multi?threshold MRI image segmentation algorithm based on Curvelet transformation and multi?objective particle swarm optimization [J]. Journal of computer applications, 2016, 36(11): 3188?3195.
[5] 李積英,黨建武,王陽萍.融合量子克隆進(jìn)化與二維Tsallis熵的醫(yī)學(xué)圖像分割算法[J].計(jì)算機(jī)輔助設(shè)計(jì)與圖形學(xué)學(xué)報(bào),2014,26(3):465?471.
LI Jiying, DANG Jianwu, WANG Yangping. Medical image segmentation algorithm based on quantum clonal evolution and two?dimensional tsallis entropy [J]. Journal of computer?aided design &; computer graphics, 2014, 26(3): 465?471.
[6] 葛雯,姬鵬沖,趙天臣.NSST域模糊邏輯的紅外與可見光圖像融合[J].激光技術(shù),2016,40(6):892?896.
GE Wen, JI Pengchong, ZHAO Tianchen. Infrared and visible light images fusion of fuzzy logic on NSST domain [J]. Laser technology, 2016, 40(6): 892?896.
[7] 徐正則.基于深度圖像動(dòng)作捕捉技術(shù)虛擬主持人的應(yīng)用研究[J].現(xiàn)代電影技術(shù),2016(8):21?26.
XU Zhengze. Application research of virtual host based on motion capture of deep image [J]. Advanced motion picture technology, 2016(8): 21?26.
[8] 王澤民.戲曲人物動(dòng)作捕捉技術(shù)的研究與應(yīng)用[J].科技創(chuàng)新導(dǎo)報(bào),2014(15):229?230.
WANG Zemin. Research and application of the motion capture technology of the characters in traditional Chinese opera [J]. Science and technology innovation herald, 2014(15): 229?230.
[9] 王新,朱信忠,趙建民,等.基于實(shí)時(shí)動(dòng)作捕捉技術(shù)在影視動(dòng)畫中的研究[J].微型電腦應(yīng)用,2014,30(3):16?17.
WANG Xin, ZHU Xinzhong, ZHAO Jianmin, et al. Research of film and animation based on the real?time motion capture technology [J]. Microcomputer applications, 2014, 30(3): 16?17.
[10] 范秀云.恐怖谷理論與動(dòng)畫電影中的逼真人物形象[J].當(dāng)代電影,2014(6):187?190.
FAN Xiuyun. Horror valley theory and realistic characters in animated movies [J]. Contemporary cinema, 2014(6): 187?190.