Estimation of disparity maps through an evolutionary algorithm and global image features as descriptors

In several approaches that include analysis processes – the most well-known being object tracking, video understanding, automatic surveillance systems, and image reconstruction – there are basic tasks to be performed. One of these tasks is related to how to select an image feature window in a frame and then compute its displacement in another frame. In the literature, the last two tasks represent an open research topic because (1) estimation of the similitude for a region window involves a set of invariants that are scene-dependent; (2) a general method for detecting the best-fitting region criterion to compute the displacement is dependent on the similarity criterion and numerical approaches for estimating the displacement; and (3) the type of conditions must be warranted so that an image feature has a high probability of estimating the displacement and numerically reaching a convergence state. In this paper, we propose a framework to estimate the displacement of an image feature from a reference image to another image. The proposal uses a generalization of the optimization concept, that is, a random search process in the dissimilarity metric space. This approach is carried out in a discrete space by mapping the variable domain to be estimated to a symbolic space with a set of operators, where a random search method is described through a uniform sampling process and genetic operators. The approach searches for the best suboptimal solution of the locality under a predefined metric criterion, avoiding divergence for the worst suboptimal solutions. The proposal is based on the formalization of the Lucas and Kanade approach. It considers as a metric-space solution the proposal of the Shi and Tomasi approach, but instead of a Taylor series expansion and step-descendant approach to solve the system, an evolutionary algorithm is used. The reference approach is well accepted as one of the most important approaches in motion displacement. To test our approach, we take the task of building a disparity map for 3D geometry extraction given the number of times the displacement computation is performed (once for each pixel). Finally, the results demonstrate that the evolutionary approach increases the repeatability and robustness of the distance estimation. [Display omitted] •A new approach to the estimation of image displacement.•An approach based on evolutionary algorithms.•An approach applied to the estimation of the Disparity Map problem from stereo vision.•Introd