The overall goal of this proposal is the high precision digital stereo obtained from a pair of images (and more generally from an arbitrary number of images of the same scene). The project goal is to reach the effective limits when we control the entire chain of acquisition to the maximum precision, with very low b/h, and with noiseless images. The proposal is a natural extension of the work of both teams in the last six years on the following topics: aerial and spatial stereoscopy, image denoising, image fusion and registration, camera calibration with high precision, cloud processing points with high fidelity to the original points. These works are based on several fundamental theorems demonstrated by the team, and never properly exploited before. They demonstrate the theoretical feasibility of a new concept of stereo allowing very high precision. For this, both teams will assemble and complete the first chain of stereo algorithms built in high accuracy ever, and demonstrate its feasibility with real cameras on real scenes. The planned generic system will use stereo cameras, placed close together, taking pictures that are almost identical, with a very low noise acquisition level and a very low aliasing, obtained by temporal image fusion. This device will be immediately feasible with market cameras, on fixed scenes for the reconstruction of urban scenes for example, or in the future with large-aperture cameras for moving scenes. Both teams were awarded recently for their achievements in stereo-vision and image denoising by international conferences and international juries. They started five years ago the project MISS (Mathematics for Stereoscopic Imaging and Space) in collaboration with CNES, which allowed them to specify two satellite projects of Earth Observation, creating stereo pairs in near-simultaneous low b / h . The first satellite, Pleiades, is currently in-orbit. The preparatory work for the operation of the Pleiades, and a fundamental theorem explained in this proposal have demonstrated that the main limit to the high precision in stereo is given by the SNR. More precisely, the theorem shows that the accuracy increases super-linearly with respect to the SNR. It is therefore very important to seize this opportunity. For civil and military applications on terrain modeling, by increasing the signal to noise ratio in an arbitrary ratio, one can reach the creation of perfect images (without aliasing). In addition the joint team showed in a preliminary draft of the ANR (Callisto) the feasibility of high precision camera calibration, therefore removing another technological obstacle hindering the transformation of stereoscopy into a high precision technique. The project’s challenge is to mount, using simple reflex good quality cameras, a complex digital processing chain, with all steps characterized in precision. The funding requested will pay two research engineers full-time over three years for the chain assembly in coordination with the two teams. This team of experienced researchers, all doctors in applied mathematics and computationally very competent, will specify and implement a generic algorithmic chain, adaptable to any device made of good quality cameras, and turning it into a digital terrain generator. The objectives are a reorganization of the research around the high precision, which will include publishing online in the new journal created by the team, IPOL (Image Processing On Line) which provides for each article online demonstrators. Contacts have already been established with two high-tech companies (DxO, Technicolor) which are exploring several stereo camera concepts. Technology transfer in the form of licensing of algorithms is expected at the end of the project.
Monsieur Jean-Michel Morel (Centre de Mathématiques et Leurs Applications) – firstname.lastname@example.org
The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.
CMLA Centre de Mathématiques et Leurs Applications
IMAGINE Laboratoire d'Informatique Gaspard Monge
Help of the ANR 299,540 euros
Beginning and duration of the scientific project: December 2012 - 36 Months