Online robot odometry calibration over multiple regions classified by floor colour

Abstract

Floor sensors allow a mobile robot to segment the environment into useful regions with properties associated with the floor, such as odometry calibration, cleaning requirements and semantic map labelling. This paper describes an accurate floor colour sensor and presents experimental results to show its effectiveness at classifying different surfaces using a Support Vector Machine (SVM). The sensor is mounted under the robot with its own light source, thus avoiding extraneous light and classifies only the floor that the robot is currently travelling over. The sensor is applied to the calibration problem of correcting systematic odometry errors of a differential drive robot. This can improve SLAM map quality by segmenting the environment into distinct regions with different odometry calibration parameters. Region based calibration of odometry is achieved using an Extended Kalman Filter (EKF) and correlative laser scan matching. This paper uses an odometry correction cost function derived from graph SLAM to show experimentally that the calibration with multiple classified regions is superior to calibration without floor classification. This paper also provides experimental results confirming that odometry calibration parameters depend on floor surface type.