Efficient and precise stereoscopic vision for humanoid robots

Abstract

This thesis is realized in the context of the RoboCup contest: a competition where two teams of robots play against each other in a soccer game. The purpose of this work is to determine if a stereo vision system could be implemented on a constrained robot platform, and provide, in real-time, useful 3D information about the playing area and the game elements. This paper starts by giving a theoretical explanation of the principles of stereo vision systems, followed by a quick review of the state of the art. As the computation power of the considered robot platform is limited, this paper then proposes an adaptation of an algorithm developed by Sudeep Pillai, claiming to achieve a good semi-dense approximation of the 3D environment at a frame rate that can reach 120Hz on a single CPU thread. The testing of this algorithm with a stereo setup of two wide-lens cameras separated by a small distance shows that the depth of a soccer ball can be estimated with a mean absolute error of 5cm/m, by directly looking at the depth of generated 3D points supposed to belong to the ball. Another analysis also reveals that the inclination at which the floor is observed by the cameras can be estimated with a precision of less than 1 degree. It is thus likely that the accuracy of the ball localization could be further improved taking advantage of this precise floor plane estimation instead of assuming that the 3D surface of a ball would always be correctly rendered by the algorithm.