Abstract
Abstract How can we localize ourselves within a building solely using visual information, i.e., when no data about prior location or movement are available? Here, we define place categorization as a set of three distinct image classification tasks for view matching, location matching, and room matching. We present a novel image descriptor built on texture statistics and dynamic image partitioning that can be used to solve all tested place classification tasks. We benchmark the descriptor by assessing performance of regularization on our own dataset as well as the established Indoor Environment under Changing conditionS dataset, which varies lighting condition, location, and viewing angle on photos taken within an office building. We show improvement on both the datasets against a number of baseline algorithms.
Highlights
Humans possess the remarkable ability to reliably localize themselves in the world under a multitude of conditions: Indoors and outdoors, in unknown terrain, with reduced senses during different weather conditions, and even under cognitive load while processing other tasks
We present a novel image descriptor built on texture statistics and dynamic image partitioning that can be used to solve all tested place classification tasks
It is important to distinguish these requirements from the requirements for vision-based Simultaneous Localization and Mapping (SLAM; see Section 1.1 for a more detailed discussion)
Summary
Humans possess the remarkable ability to reliably localize themselves in the world under a multitude of conditions: Indoors and outdoors, in unknown terrain, with reduced senses during different weather conditions, and even under cognitive load while processing other tasks. The features required there (tracking features) need to be stable across successive camera frames and recognizable from slightly different viewpoints, whereas the features suitable for vision-only place recognition have much stronger invariance requirements. For indoor localization [4], use histograms over relatively simple orientation descriptors on the Indoor Environment under Changing conditionS (INDECS [5]) database. These features are very generic, and confusion may arise if the same texture feature is found in multiple areas of an image, e.g., at the ceiling and on the floor. We present an approach that circumvents this problem by partitioning the image according to texture occurrences
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
