A kinect-based SLAM in an unknown environment using geometric features

Abstract

This paper proposes a geometric feature-based method to solve the Simultaneous Localization and Mapping (SLAM) problem in an unknown structured environment using a short range and low Field of View (FoV) measurement unit such as Kinect sensor. A RANdom SAmple Consensus (RANSAC) based algorithm is used for feature detection, and a grid-based point cloud segmentation method has been introduced to improve the multiple feature point-detection in a 2D depth frame. A fast SLAM algorithm is used to estimate the robot posterior and the map of the environment. This approach builds the individual maps for each particle using geometric features that are extracted from a 2D slice of a 3D depth image. Each map contains individual Extended Kalman Filters (EKFs) for each and every feature-point. This method reduces the uncertainty of the robot pose in the prediction step and it improves the pose accuracy when more geometric feature-points are available. The proposed feature-based approach gives better localization and compact map representation in structured environments when distinct features are available. The importance weighting and the comparison of features with the on-line map are performed according to the maximum likelihood criterion. In order to reduce the particle depletion, the map is updated only when a new Odometry measurement and new range measurements are available. The experiments are carried out using the recorded data with a non-holomonic mobile robot equipped with a Kinect sensor in a small scale indoor structured environment. For comparison, the grid based SLAM result is also presented for the same data set.