Accurate polygon-based 3D SLAM with encoder-edge based state estimation and dynamic scan error compensation

Abstract

Localization and map building is essential parts in autonomous mobile robot navigation. Localization determines where mobile robot is in a given environment. Map building constructs the environment map with assumed structures. In general, localization and map building are performed simultaneously so called Simultaneous Localization and Mapping (SLAM). Task completion rates of mobile robot is determined by how the mobile robot reaches where has to accomplish a task well. Even though the mobile robot is autonomous and performs tasks by itself, user finally determines the goal position of mobile robot and establishes strategy how mobile robot accomplishes tasks. Therefore, the accuracy of SLAM is very important. Furthermore, 3D map is natively suitable to our 3D space and has more useful information. Therefore, accurate 3D SLAM is a challengeable and important problem in robotics.

In this dissertation, our objective is to establish accurate 3D SLAM. Typical SLAM is composed with five major parts: dead-reckoning, feature extraction, matching, pose estimation, and map building. Improved performance of each major part can induce accurate SLAM. This dissertation approaches to achieve the accurate 3D SLAM by improved the performance of some major parts. Among five major parts, we select three parts: dead-reckoning, feature extraction, and map buildling. All major parts are important for accurate 3D SLAM, but selected three parts have influences on other major parts, one another, and the overall performance of 3D SLAM effectively.

Firstly, we propose Encoder-Edge based State Estimation (EESE). Conventional periodic encoder sensing is used to obtain the quantized motor position. But we performed detection of the encoder edge’s time and value precisely. The EESE uses the edge-time Kalman filter (ETKF) which performs predictions at edge time and periodic sampling time, and update at edge time. Since encoder edges are detected too often at high motor speed or high encoder resolution, only the first edge in each sampling interval is utilized to reduce the computation time. The EESE guarantees far more accurate state estimation with even low encoder resolution and uncertainty on motor parameters. Performance of the EESE is validated through simulations and experiment about single motor first. And then, we applied the EESE to a two-wheeled mobile robot (TMR) for more practical situation. We confirmed the performance of EESE through the motion of TMR with and without EESE.

Next, we propose Dynamic Scan Error Compensation (DSEC) for LRF. The accuracy of conventional LRF is high, but it is not true when robot is moving. Since one LRF scanning time is very small and it is not zero, all scan ray times are different. If the time difference between scan rays is not considered during scanning, all scan rays are distorted. The distortion is critical error of LRF and can be called as dynamic scan error. This impairs the accuracy of extracted features. The source of dynamic scan error is that LRF can not know where and when scan rays are measured. In order to compensate the dynamic scan error, positions of the center of LRF every scan rays have to be obtained. From encoder data, we can obtained the positions at sampling times. Since the time invertal between scan rays is smaller than encoder sampling time, we use interpolation to compute the positions. DSEC is performed with transformation matrices with simplification to reduce the computation time because about hundreds of scan rays have to be compendated every scanning. The DSEC guarantees far more accurate LRF scan measurement at even robot motion with variable speed. Performance of the DSEC is validated through simulations of DSEC itself and 2D SLAM with and without DSEC.

Finally, we propose 3D map building and 3D SLAM based on 3D polygon map. 3D SLAM has been studied to provide more information about environment while it needs far more data processing and computation loads. For less data load, 3D infinite plane is mostly used in recent 3D SLAM. But it is not appropriate to describe 3D envrionment. We propose 3D polygon map with vertices which restricts the area of 3D plane. 3D polygon map with vertices is more less data load and useful to describe 3D environment. In order to reduce computation time and enhance the accuracy of map building, we propose consecutive sequential constructions to build 3D polygon map. First, 3D scan vector sequential bi-directional extraction will be performed. When some 3D scan vectors are collected, 3D infinite plane is extracted sequentially. From grouped 3D scan vectors in 3D plane, 3D edge vectors sequential extraction is performed. Using 3D edge vectors, we extract vertices for 3D polygon map. Proposed 3D map building is performed well in several indoor environments and the potential of that is validated in outdoor environment. Hybrid 3D SLAM of 2D SLAM method and proposed algorithm is proposed. For accurate 3D SLAM, we integrate EESE and DSEC with 3D polygon map. Using the integration, accurate polygon based 3D SLAM with EESE and DSEC is validated by experiments.