1P1-W06 部品SLAM : 教師なし部品モデルに基づく簡潔・識別的な情景マッチング手法(ロボットビジョン)

Abstract

We consider the task of long-term visual SLAM, i.e., simultaneous localization and mapping, in a partially changing environment (SLAM-PCE). The main problem we face is how to obtain discriminative and compact visual landmarks, which are necessary to cope with changes in appearance in an environment and with a large amount of visual information. We address this issue by proposing the use of common object patterns, which are inherent in typical environments (e.g., indoor, street, forests, suburban, etc.), as visual landmarks for a SLAM-PCE task. In our contributions, we describe our approach, "part-based SLAM", and validate its effectiveness within a standard problem of view image retrieval. The main novelty of this approach lies in that the common landmark objects are extracted in an unsupervised manner via common pattern discovery, and can be used for compact characterization and efficient retrieval of view images. Our method is also innovative in its use of traditional bounding box-based part annotation: an image is represented in a compact form, "bag-of-bounding-boxes (BoBB)"; and then, the scene matching can be solved efficiently as a low dimensional problem of matching bounding boxes. The results of challenging experiments show that it is possible to have high retrieval performance with compact image representation with only 16 words per image.