Associating Locations Between Indoor Journeys from Wearable Cameras

Abstract

The main question we address is whether it is possible to crowdsource navigational data in the form of video sequences captured from wearable cameras. Without using geometric inference techniques (such as SLAM), we test video data for its location-discrimination content. Tracking algorithms do not form part of this assessment, because our goal is to compare different visual descriptors for the purpose of location inference in highly ambiguous indoor environments. The testing of these descriptors, and different encoding methods, is performed by measuring the positional error inferred during one journey with respect to other journeys along the same approximate path. There are three main contributions described in this paper. First, we compare different techniques for visual feature extraction with the aim of associating locations between different journeys along roughly the same physical route. Secondly, we suggest measuring the quality of position inference relative to multiple passes through the same route by introducing a positional estimate of ground truth that is determined with modified surveying instrumentation. Finally, we contribute a database of nearly 100,000 frames with this positional ground-truth. More than 3 km worth of indoor journeys with a hand-held device (Nexus 4) and a wearable device (Google Glass) are included in this dataset.