Weighted online calibration for odometry of mobile robots

摘要

Accurate odometry is required for robust and reliable localization of mobile wheeled robots. Therefore, odometry is combined often with a positioning system and data fusion algorithms. However, odometry parameters are not known in advance and vary depending on environment and age of the vehicle. A typically approach to solve this problem is to employ on-demand calibration. Current approaches are sensitive to measurement noise and shape of the trajectory followed by the robot. Furthermore, a common approach is to integrate calibration into localization and mapping, which makes it hard to reason about the behavior due to feedback. We present an approach that describes the calibration as an error minimizing curve fitting problem. Position measurements are pre-filtered with a tracking filter to reduce noise and to generate weights. Weighted non-linear least squares is used to calculate optimal odometry parameters for a given trajectory. The trajectory is heuristically weighted and an exponential smoothing filter operates on the resulting odometry parameters to track a vector of odometry parameters. The optimization is continuously executed quasi-online with partial trajectories. We simulate multiple runs of the algorithm on different trajectories in a simulated environment. We validate the approach with a Roomba robot and two different positioning systems, a UWB based method and an optical landmark based method. We discuss the results, general convergence characteristics and stability.