manuscript No EXIF D to appear in Experiments in Fluids
15
pages
English
Documents
Obtenez un accès à la bibliothèque pour le consulter en ligne En savoir plus
Découvre YouScribe et accède à tout notre catalogue !
Découvre YouScribe et accède à tout notre catalogue !
15
pages
English
Documents
Obtenez un accès à la bibliothèque pour le consulter en ligne En savoir plus
manuscript No. EXIF-D-10-00011 to appear in Experiments in Fluids Fast and accurate PIV computation using highly parallel iterative correlation maximization F. Champagnat1, A. Plyer1, G. Le Besnerais1, B. Leclaire2, S. Davoust2 and Y. Le Sant2 the date of receipt and acceptance should be inserted later Abstract Our contribution deals with fast computa- tion of dense two-component (2C) PIV vector fields us- ing Graphics Processing Units (GPUs). We show that iterative gradient-based cross-correlation optimization is an accurate and efficient alternative to multi-pass processing with FFT-based cross-correlation. Density is meant here from the sampling point of view (we ob- tain one vector per pixel), since the presented algo- rithm, folki, naturally performs fast correlation op- timization over interrogation windows with maximal overlap. The processing of 5 image pairs (1376 ? 1040 each) is achieved in less than a second on a NVIDIA Tesla C1060 GPU. Various tests on synthetic and ex- perimental images, including a dataset of the 2nd PIV- Challenge, show that the accuracy of folki is found comparable to that of state-of-the-art FFT-based com- mercial softwares, while beeing 50 times faster. 1 Introduction Particle Image Velocimetry (PIV) has become an essen- tial tool for flow diagnosis and is therefore widely used in industrial as well as academic situations.
Voir
- based
- pixel
- based cross-correlation
- piv
- spatial resolution
- lk techniques
- using gpu versus cpu
- very robust
- gpu
manuscript No. EXIF-D-10-00011 to appear in Experiments in Fluids
Fast and accurate PIV computation using highly parallel iterative correlation maximization F. Champagnat 1 , A. Plyer 1 , G. Le Besnerais 1 , B. Leclaire 2 , S. Davoust 2 and Y. Le Sant 2
the date of receipt and acceptance should be inserted later
Abstract Our contribution deals with fast computa-tion of dense two-component (2C) PIV vector fields us-ing Graphics Processing Units (GPUs). We show that iterative gradient-based cross-correlation optimization is an accurate and efficient alternative to multi-pass processing with FFT-based cross-correlation. Density is meant here from the sampling point of view (we ob-tain one vector per pixel), since the presented algo-rithm, folki , naturally performs fast correlation op-timization over interrogation windows with maximal overlap. The processing of 5 image pairs (1376 × 1040 each) is achieved in less than a second on a NVIDIA Tesla C1060 GPU. Various tests on synthetic and ex-perimental images, including a dataset of the 2nd PIV-Challenge, show that the accuracy of folki is found comparable to that of state-of-the-art FFT-based com-mercial softwares, while beeing 50 times faster. 1 Introduction Particle Image Velocimetry (PIV) has become an essen-tial tool for flow diagnosis and is therefore widely used in industrial as well as academic situations. Its current limitation is however the time necessary to compute the vector fields from the images, which often imposes spe-cific constraints in the schedule of test campaigns. In that respect, the important development of high speed PIV systems over the last decade appears even more challenging. We propose a solution to shorten dramat-ically this processing time, based on an algorithm that 1 Information Processing and Modelling Department, French Aerospace Lab (ONERA), 29 avenue de la Division Leclerc, 92322 Chatillon Cedex, France fchamp@onera.fr 2 Fundamental and Ex-perimental Aerodynamics Department, French Aerospace Lab (ONERA), 8 rue des Vertugadins, 92190 Meudon, France
computes dense 2C vector fields using Graphics Pro-cessing Units (GPUs). GPU has already been compared to other architec-tures for PIV processing in previous works ( Schiwietz and Westermann , 2004 ; Venugopal et al , 2009 ). These studies concentrated on cross-correlation using FFT, but the speed-up factor for FFT using GPU versus CPU architecture does not exceed three. In this con-text, real-time computation therefore requires large PC clusters with a GPU at each node ( Venugopal et al , 2009 ). Former real-time realizations also involve paral-lelisation on Field-Programmable Gate Arrays (FPGA) ( Iriarte Munoz et al , 2009 ; Yu et al , 2006 ). Although efficient and convenient for embedded systems, this so-lution is far more expensive than GPU to implement, both in terms of hardware cost and of software develop-ment effort. Interestingly, these architectures get rid of FFT in favor of direct correlation, which is better suited to FPGA architectures. In contrast to these works, the approach proposed hereafter relies on a technique for cross-correlation maximization that departs from the classical FFT method, or from direct correlation. Its structure is ideally matched to massively parallel archi-tectures, and therefore allows a 50 times speed-up using a single GPU. The algorithm folki (French acronym for Iterative Lucas-Kanade Optical Flow, Le Besnerais and Cham-pagnat , 2005 ) was originally designed in the context of computer vision for motion estimation in video se-quences. But folki proved also very robust and adap-tive to many other kinds of images such as those ob-tained in photomechanics and PIV. It is based on the classical interrogation window paradigm, but belongs to the family of Lucas-Kanade (LK) algorithms (see Baker and Matthews , 2004 , for a review). The basic LK method is already known in the PIV community
