Recursive estimation of three-dimensional unmanned aerial vehicle motion and structure based on the L ∞ -norm

This article presents a non-linear filtering framework to estimate three-dimensional motion and structure from a stream of images acquired by a monocular camera. Interest in this article is related more specifically to the case where the camera is onboard an unmanned aerial vehicle. Although there have been numerous attempts to solve the problem stated above, most formulations were based on the L₂-norm, mainly represented by the extended Kalman filter (EKF). Instead, a new approach based on the L_∞-norm minimization criterion is developed. Unlike former methods, framework described in this article allows better treatment of the linearization of the perspective projection function that is associated with inherent errors in the EKF solution and can lead to severely degraded performance. Moreover, the L_∞ criterion does not assume Gaussianity or whiteness of the noises providing an elegant solution for the problem of biased measurements. Results on synthetic and real aerial imagery demonstrate the findings.