Vision-based object search in unknown human environment using object co-occurrence graph

Object search in unknown human environment by utilizing Object Co-occurrence Graph (OCG) is presented. In our approach, OCG which represents co-occurrence relations between objects is employed as robot's knowledge. Since direct search that solely relies on the finding of target object in each scene may face some difficulties such as insufficient image's resolution, illumination change, or occlusion problem, indirect search utilizing hints from large-size objects that relate to a target object may provide better solutions and can make the task become easier and faster. In this work, OCG is automatically constructed based on word-tagging images provided by millions of Internet users. Moreover, in order to illustrate the applicability of OCG, practical robot's experiment that performs object search autonomously without prior information of environment's layout and object locations is demonstrated. In addition, experimental result supports our assumption about the simplicity of the finding of target object, especially small-size object, in unknown environment when utilizing information from clue objects.