서비스 로봇을 위한 의미적 지도작성과 능동적 자기위치추정

Abstract

In general, a semantic map is created after a metric map is built with accurate sensors. However, technology is needed to allow a semantic map to cooperate with humans and to enable use of this map with a low-grade sensor. This dissertation proposes a technology to build a semantic map using a low-grade sensor. To overcome problems related to semantic mapping with a low-grade sensor, I have developed a human-navigation-inspired, semantic map-building, active semantic localization, and probabilistic Bayesian model for service robots. Four human navigation strategies (path integration, view-dependent place recognition, reorientation, and active searching for additional landmarks) are integrated in proposed method as a global-topological map, egocentric local-metric map, reorientation rules, and active semantic localization capability, respectively. I developed a Bayesian model to support my proposed semantic representation. This model is validated in such way that map-building and localization accuracy improve whenever objects and their spatial relationships are detected and instantiated. Although the localization accuracy is relatively poor, a service robot can build a semantic map and easily localize itself through reasoning, which is not a one-step action resulting from Markovian localization but a subgoal to find sufficient evidence. Additionally, the semantic map can be one world model by which service robots can infer knowledge. Experiments were performed with a single camera, and the Evolutionary Robotics Software Package (ERSP) commercial software was used to recognize surrounding objects. The distance and bearing measured using the single camera and ERSP were often inaccurate. Nevertheless, I successfully created a semantic map and estimated the location of the robot by changing inaccurate distance and bearing to approximate metrics using the proposed method. Semantic map building took place in a 14x26.5 m corridor and a 7x8 m room. The proposed method enables a robot to navigate effectively in a service environment using navigation schemes inspired by humans and their actions.