Position rectification with depth camera to improve odometry-based localization

Abstract

Indoor localization plays an important role for many applications especially robotics where the location of robot is necessary for the tasks of tracking and controlling. Among efforts proposed to address this problem, odometry-based localization is presented as an effective method with simple installation. This approach is based on the movement information of the robot wheels that are obtained by motion sensors to estimate position changes of robots. However, the errors of estimating the poses of robot by odometry are accumulated and increases overtime due to the wheel skidding, sensor errors, or inaccurate information of wheelchair configuration. This paper presents an approach to overcome these drawbacks through the tuning process using a depth camera. A randomized forest is trained to map depth images captured by a depth camera to labeled locations. As a sequence, this information is used to refine the results of odometry-based localization. Experiments are conducted with a real electronic wheelchair to show that the proposed approach helps increase the accuracy of the conventional odometry-based methods. Therefore, the approach is applicable for many applications related to indoor localization.