A Generic Road-Following Framework for Detecting Markings and Objects in Satellite Imagery

In order to construct algorithmic solutions to the problem of geolocalization of user-generated photographs and videos, one must first populate a geographic database with the different types of objects and markings that are likely to be seen in the photographs and videos. Toward that end, this paper presents a framework for detecting and labeling objects in satellite imagery. The objects that we are interested in are characterized by low-level features that exist mostly at or beyond the limits of spatial resolution in the satellite images. To deal with the challenges posed by the extraction of such features from satellite imagery, we confine our search to the vicinity of the OpenStreetMap (OSM) delineated roads in a geographic area. The OSM roads are projected into the satellite images through inverse orthorectification. As an illustration of the performance of the object detection framework presented in this paper, our system can detect pedestrian crosswalks in a 200000 sq. km. region of Australia (that is covered by 222 satellite images) with a recall rate of 63% and a precision of 89% (evaluated on a 100 sq. km. subregion). All of the computer processing for this result takes a total of 6 h on our in-house cloud computing framework that consists of five nodes, each an off-the-shelf high-end PC class workstation.