In this work we study the gain offered by massively parallel architecture for low-level vision processing. Reviewing the recent evolutions of computer architectures, we underline the massively parallel solutions which are now available to every programmer, the GPU. Exploiting the computing power of these solutions requires a change in programming methodology. However, as far as low-level vision is concerned, we show that most problems can be solved using a restricted number of programming patterns. Finally, we propose a new model to evaluate the complexity of solutions developed on GPUs. Then we apply these programming principles to various low-level vision problems. The first one is optical flow (OF) estimation, which is at the core of many video processing problems. We present an OF estimation on GPU, called FOLKI, which allows good estimation quality on various real-world video sequences with unprecedented computing performance. An application of this work concerns particle image velocimetry, a measurement technique of paramount importance for experimental fluid mechanics. Then we turn toward super-resolution (SR). We first introduce an original and very fast solution, which uses the FOLKI optical flow estimate for image registration. Then various solutions of increasing computational complexity are proposed, which lead to better resolution and robustness. Using these techniques we show very original SR results on video with complex motions (walking person, moving vehicle). Finally the last chapter shows on-going work on 3D measurement contexts in experimental physics and robotics.
