Extracting Human Mobility Data from Geo-tagged Photos

Photos shared by users on public websites (Flickr, Panoramio) provide a resource for mining behavioural data. When the photos are associated with locations and time stamps, we can reconstruct the trajectories of the users and use the resulting mobility traces for learning behaviour patterns. In this paper we focus on two aspects of mobility traces: noise filtering and semantic annotation. The extracted trajectories are initially noisy due to errors in geographical coordinates and time stamps. We show how such noise can be partially filtered and evaluate the performance of the filtering on a synthetic dataset. To make use of the mobility traces, an essential step is semantic annotation. Places or activities are associated with segments of the traces. This is frequently performed by integrating a database of relevant places and associating them by proximity. We demonstrate that the popularity of the places, if available, can improve the association accuracy. In our experiment, the accuracy of automatic annotation increases from 60% to 68%.