Rapid Obtaining of Image Matching for Robotic Vision

Abstract

Robotic vision gives robots the ability to interpret the scene by emulating human vision. This can be achieved with a pair of cameras that obtain images of the environment from two points of view. A three-dimensional model is obtained, in units of length, by means of a stereoscopic technique. For its calculation, it is necessary to know the physical characteristics of the image acquisition system, which are generally obtained only once when the system is calibrated, at the beginning. Then, for the three-dimensional reconstruction it is essential to have the coordinates of points in both images that correspond to each point in the scene. This process is called recognition of
homologous points, correspondence or aspect and has a very high computational complexity. This work focuses on this task. It is necessary to automatically recognize that they are present in both images to achieve the association of points in the scene. To do this, edge analysis of objects in images is a common approach. One problem that immediately arises is that the shapes are in any position, size, and orientation within the images. A descriptor pattern of edge curves is presented that is invariant against rotations, translations and scaling, consisting of a discrete curve. Achieving the recognition of an object in an image then consists in measuring the distance between the descriptor pattern of the object to be searched for and the patterns of each object in the image and comparing their difference with a predefined threshold of similarity. This pattern
is dependent on the determination of a starting point. However, after obtaining the orientation of the shape and given that in stereoscopy, both images are normally similar, so are the objects contained therein. Then the identification of the starting point is easily achieved. With this,making the multi-point appearance of an object’s edge curves can be achieved with less difficulty,
allowing for an approximate three-dimensional representation that can be compatible with realtime requirements. Experimental data is included