Ceiling Vision-Based Active SLAM Framework for Dynamic and Wide-Open Environments

Abstract

Since most successful navigation tasks of robots heavily rely on the smooth operations of three functions

mapping, localization, and exploration, they have been crucial issues for practical deployment of the service robot in indoor public spaces during the last two decades. Especially, Simultaneous Localization and Mapping (SLAM) problem is to estimate a pose of a robot and a map simultaneously while the robot explores in unknown environments.The typical indoor environment can be divided into three categories

office (or room), hallway, and wide-open space such as lobby and hall. There have been numerous approaches for solving SLAM problem in office (or room) and hallway. However, direct application of the existing approaches to wide-open space may be failed, because it has some distinguished features compared to other places as follows.1. It is difficult to detect objects using a laser scanner with 4 m detection range 2. It has high ceiling3. There are many moving persons and dynamic objects4. It frequently changes by the installation of unexpected inner structuresTo solve this problem, this thesis proposes a new ceiling vision-based active SLAM framework, with an emphasis on practical deployment of service robot for commercial use in dynamically changing and wide-open environments by adopting the ceiling vision to those challenging environment as follows. First, we introduce a new visual node descriptor (VND) which is a set of edge points and their orientations in image space. It is a model-free landmark, and thus can be extracted from ceiling image regardless of complexity of ceiling pattern. Since the VND consists of edge points, we use Iterative Closest Point algorithm, which is known to be robust to partial occlusion and small noise, for matching between two VNDs (i.e. data association). Also, we can reduce false data association through visual servoing of the robot, iteratively verifying VND matching. Lastly, the proposed algorithm does not need to estimate the height of visual feature, therefore it is irrelevant to ceiling height.Second, a recursive ‘explore and exploit’ is proposed for autonomous mapping. It is recursively performed by spreading out mapped area gradually while the robot is actively localized in the map. It can improve accuracy of map due to occurrence of frequent small loop closing.Third, dynamic edge link (DEL) is proposed to represent the effect on environmental changes in the map. We do not need to filter out corrupted sensor data and distinguish moving object from static one by using DEL. Also, self-repairing map mechanism is introduced to cope with unexpected installation or removal of inner structures. The environmental changes are incorporated into the DEL in online manner, thus the robot can figure out which part of the map should be reconstructed using the DEL. We therefore achieve long-term navigation due to self-repairing map.Several simulations and real experiments in various places show that the proposed active SLAM framework could build a topologically consistent map with average maximum error of 0.4 m, and demonstrated that it can be applied well to real environments such as wide-open space in a city hall and railway station.