Estimación de pose por filtrado complementario de marcadores y odometría

Abstract

Accurate estimation of rigid object pose is a fundamental requirement in vision-based robotic. In eye-in-hand configurations, estimates obtained through Perspective-n-Point (PnP) from visual markers exhibit frame-to-frame noise.  We propose a recursive method that combines visual observations of ArUco markers with the robot odometry. The odometry transform the observations into the robot base coordinate frame, where the object remains static, and a robust exponential weighted moving average (EWMA) formulated on the Lie group SE(3) is applied to them. The method incorporates a state-dependent Huber weighting term and a reprojection error-based weight from PnP to model the reliability of each observation. Experimental validation is carried out using a robotic arm with a monocular camera observing a static object with ArUco markers. The results show a significant reduction in relative pose error and temporal jitter compared to direct PnP and EWMA variants without robust weighting.