Performance evaluation of visual odometry using an industrial robot as ground truth

Computer vision algorithms for object localization and tracking are largely used in industry and academic research. Nonetheless, the performance evaluation of these algorithms is still an open problem, because the common systems used as ground truth are expensive and often not available in industry or research labs. In this paper we propose a method to evaluate the performance of a visual odometry algorithm using an industrial robot. The parameters required for the comparison are estimated by reducing the problem to a Quadratically Constrained Quadratic Program. The method is applied to an experimental set up composed of a six degrees of freedom industrial robot and a commercial monocular camera.