Drone mapping of damage information in GPS-Denied disaster sites

Abstract

The increasing number and severity of natural hazard events in recent years, and their devastating impact on human life, local economies, and the built environment has called governments around the world into action and created a new mandate for a paradigm shift in disaster management and mitigation policies. To this end, new affordable technologies with mobile connectivity (e.g., smartphones, unmanned systems, reality capture devices) have scaled up tasks such as data collection and curation, leading to a significant increase in the volume of data gathered and shared in the aftermath of disasters. In the meantime, advancements in high-power and distributed computing have created new opportunities in fast and reliable data analytics. In particular to the application of drones in disaster response, past research has primarily focused on aerial data collection and more recently, ground object detection. Geolocalization of drone data (i.e., the process of determining the geographical position of objects in drone’s field of view), however, is a complex task that relies on prior knowledge of the drone’s geolocation (e.g., flight path coordinates, inertial sensors, camera gaze). Such metadata may not be always available or shared across platforms especially with the increased use of crowdsourcing in disaster response, damage assessment, and recovery. This paper presents a methodology for spatial mapping of disaster impact information in drone videos without reliance on GPS data of the aerial camera. We perform progressive mapping using scale-invariant visual features in red–greenblue (RGB) videos of disaster-affected sites in two major hurricanes in North America, namely Harvey (2017) and Dorian (2019). Results indicate that the proposed methodology can project objects from the perspective view of a drone camera onto an orthogonal map with 32.7–36.9 ft of average root mean square (RMS) error in a land area of 18–45 acres.