2004, Volume 6, Issue 12
Strategic Study of CAE >> 2004, Volume 6, Issue 12
The Improving Characteristics of the Gradual Gaussian Multidimensional Pre-filter for Optical Flow Estimation
Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an 710049, China
Next Previous
Abstract
Based on the unified estimation-theoretic framework, an effective method of using the gradual Gaussian multidimensional pre-filter to improve the optical flow estimation is presented. The pre-filtering and smoothing effect, which attenuate the temporal aliasing and the interesting signal structure of the optical flow field, are altered with adjusting the spatiotemporal standard deviation parameters. The first 50 frames of the standard Flower Garden and Football video sequence are tested as the reference image sequences, and the LK algorithm as the reference optical flow computing method. Experimental results in objective evaluation show that the optimum temporal standard deviation parameter is 0.4, the optimum spatial standard deviation parameter is in a range of 1.6～2.0 under the condition that the pre-filtering window size is 5 × 5 pixels. After pre-filtering the image sequence by the Gaussian multidimensional filter, the average PSNR of the reconstructed frames enhance 2.572 dB, higher than that using the standard optical flow computing method by nearly 13.6 % .
Figures
图1
图2
图3
图4
图5
图6
图7
References
[ 1 ] 泰卡尔普AM 数字视频处理 (DigitalVideo Processing) [M ].崔之祜, 江春, 陈丽鑫译.北京:电子工业出版社, 1998
[ 2 ] BarronJL , FleetDJ, BeaucheminSS .Performanceofopticalflowtechniques[J].InternationalJournalofComputerVision, 1994, 12 (1) :43～77 link1
[ 3 ] SinghA , AllenP .Image flowcomputation:anestimation theoreticframeworkandaunifiedperspective[J].CVGIP :ImageUnderstanding, 1992, 56:152～177
[ 4 ] LucasBD , KanadeT .Aniterativeimageregistrationtechniquewithapplicationtostereovision[A ], Proceedingsofthe7thInternationalJointConferenceonArtificialIntelligence[C].Vancouver, 1981.674～679
[ 5 ] 付 昀.线性人脸对象类模型的匹配提升技术[R].西安:西安交通大学人工智能与机器人研究所, 2003
[ 6 ] NamJY , SeoJS , KwakJS , etal, Newfast searchalgorithmforblockmatchingmotionestimationusingtemporalandspatialcorrelationofmotionvector[J].IEEETransactiononConsumerElectronics, 2000, 46 (4) :934～942
京公网安备 11010502051620号
