Generación de regiones de interés con potencial de contener peatones mediante búsqueda focalizada usando visión monocular

Abstract

This article presents an algorithm for regions of interest generation with high potential to contain a pedestrians over monocular images. The generation of these regions has been built using an algorithm to generate search hyperplanes in function of the road-surface together with the generation of random windows on this area plus a variation of the pyramidal sliding window technique; then the pre-processing is done using vertical and horizontal gradient filters. To verify that the region is a possible pedestrian, we start from two hypotheses regarding the human figure, the vertical component is greater than the horizontal and the strong vertical human-symmetry. Through this process, a reduced and optimal set of Regions is obtained in the range of between 2 and 25 meters in front of the camera. The experimental results, over the state of the art databases, show that there is a 91% rate of valid windows with respect to the total of true windows, at 25.38 frames per second.