Section: New Results

Parameter estimation

Participants : Frédéric Sur, Erwan Kerrien, Raffaella Trivisonne.

Metrological performance assessment in experimental mechanics

A problem of interest in experimental solid mechanics is to estimate displacement and strain maps on the surface of a specimen subjected to a load or a tensile test. Two contactless approaches are available in the literature, based on depositing on the surface of the specimen either a pseudo-periodic grid or a random speckle. Analyzing images taken before and after deformation permits to estimate strain maps. While periodicity permits to make use of Fourier analysis in the first case, digital image correlation (DIC) is used in the second case. Concerning pseudo-periodic grids, we have investigated noise reduction techniques. While averaging a series of images is certainly the most basic option to reduce the noise, it is impossible to get rid of residual vibrations carried for instance by concrete floor slabs. We have shown in [16] that, while these vibrations indeed blur grid images, they still permit to reduce the noise amplitude in the displacement and strain maps. Concerning DIC-based techniques, we have investigated the effect of sensor noise on the measurement resolution. Since displacement of interest are most of the time below one pixel, interpolation plays a major role. We have proposed a new resolution formula which takes interpolation into account. Besides, this formula has been the subject of an experimental assessment on real data, in the presence of signal-dependent noise [24] , [18] .

Sensor noise measurement

We have investigated in [15] (additional information available in [27] ) the problem of sensor parameter estimation from a series of images, under illumination flickering and vibrations. Illumination flickering is indeed a natural assumption for indoor artificial lights. It is also involved by slight variations in the opening time of a mechanical shutter. We have proposed a model of the pixel intensity based on a Cox process, together with an algorithm which, taking benefit of flickering, gives an estimation of every sensor parameter, namely the gain, the readout noise, and the offset.

Image driven simulation

The IDeaS ANR project targets image-driven simulation, applied to interventional neuroradiology: a coupled system of interactive computer-based simulation (interventional devices in blood vessels) and on-line medical image acquisitions (X-ray fluoroscopy). The main idea is to use the live X-ray images as references to continuously refine the parameters used to simulate the blood vessels and the interventional devices (micro-guide, micro-catheter, coil). Our guideline is to follow a sequential statistical filtering approach to fuse such heterogeneous data.

Christo Gnonnou was hired as an engineer (located at Inria Lille in Defrost team (ex-Shacra), contract started on January 1st and ended on October 31st). He continued the work to specify which parameters the simulation is sensitive to, in a view to design a reduced parametric model of the device, and associate covariances to its parameters. He also worked on inverting the mechanical parameters of any device, using our experimental setup based on a high speed stereo rig.

Maxime Malgras worked on his Master's thesis in the team. His investigations aimed at designing a particle filter where the transition function is approximated by a polynomial chaos (PC) instead of particles. It appeared that PC is adapted to compute the posterior in a particle filter but it is not clear whether the number of samples required to estimate the PC coefficients is smaller than the number of particles required for the filter to be accurate, which questions the capacity of PC to reduce the computational burden of particle filters in high dimensions. Raffaella Trivisonne started her PhD thesis in November (co-supervised by Stéphane Cotin, from MIMESIS team in Strasbourg) to investigate deeper on this subject and apply data assimilation to image driven simulation of endovascular interventions.