Fast, succinct, and indirect map matching via known reference maps as intermediary: A randomized visual phrase approach

Map matching is a fundamental task in many robot vision applications, including viewpoint localization, change detection, alignment, merging, segmentation of maps, and multi-robot mapping. Existing frameworks so far have concentrated on local feature-based approach, where discriminative local features are extracted from the maps and visual indexing and map database searched are performed to find correspondence between an input query and a collection of global maps. To improve compactness and discriminativity of map matching algorithms, in this paper, we propose a novel map matching framework, “unsupervised part-based scene modeling”. Our research in this paper is motivated by two independent techniques derived from the field of computer vision: part model and common pattern discovery. First, our basic observation is that a part model is a powerful discriminative model, and it can be compact if a scene is explained by fewer larger parts. Second, in contrast to existing supervised part models that rely on pre-trained part detectors, our CPD-based approach learns a compact part model with fewer larger parts in an unsupervised manner, which enables a robot to learn a compact map model without human intervention. In our contributions, we present a practical implementation using shape context descriptors, by employing the recently developed CPD technique, randomized visual phrase. The results of challenging experiments using publicly available radish dataset show the effectiveness of our approach in terms of map matching performance, speed, and compactness of the map representation.