Unmanned drones aren’t just for warfare. In recent years, they’ve been used to map wildlife and monitor crop growth. But current software can’t always handle the vast volume of images they gather. Now, researchers have developed an algorithm that will allow drones to 3D-map scores of hectares of land in less than a day—an advance that is important for cost-effective farming, disaster relief, and surveillance operations.
“It is revolutionary for the problem of mosaicking large volumes of imagery,” says computer scientist Dalton Rosario of the U.S. Army Research Laboratory in Adelphi, Maryland, who was not involved with the study.
Camera-equipped, autonomous, unmanned aerial vehicles (UAVs) can fly low to the ground and take high-resolution images of crops that tell farmers exactly where to plant their seeds or add fertilizers—at a tenth the cost of flying a plane or purchasing satellite images. To stitch the photos together into a mosaic, a computer program needs to figure out the exact angle and position of the camera for each picture taken in order to build a 3D model of the land. Conventional software does that by looking at common features in neighboring photos—for example, the same corn plant that appears in two images—and marking them with points called tie points. The software then tweaks its calculation of the camera positions for all the photos at once, so that when it projects the tie points onto a 3D model, points from different images match up to form a coherent projection of the corn plant. This method works well for a few hundred photos, but once the number of images exceed a thousand—typical for mapping a 40-hectare farm—the process can take 1000 hours, an impossible load for desktop computers.
So computer scientist Mark Pritt and colleagues at Lockheed Martin in Gaithersburg, Maryland, took a different route. Their computer program directly projects the points from each photo onto a 3D space without knowing the exact shape of the land or the camera positions. As a result, the tie points don’t necessarily match up, which means the same corn plant can have two projections on the model. When that happens, the algorithm automatically takes the middle point between the two projections as the more accurate location and adjusts the camera position accordingly, one image at a time. Because the algorithm tweaks far fewer things at each step, the shortcut drastically speeds up calculations. Once the software has adjusted the camera positions for all the photos, the software repeats the entire process—starting from projecting the points to the 3D space—to correct for any errors.
With the new algorithm, the researchers can produce a map from a thousand images in just 4 hours, they reported this month at the annual IEEE Applied Imagery Pattern Recognition Workshop. That means it can render a map of the land within 24 hours after the drones fly, giving farmers a head start on taking care of their crops and enabling them to use drones routinely to monitor crop health.
The image-mosaicking algorithm can also speed up applications of drone imaging such as surveillance and disaster relief, says computer scientist Kannappan Palaniappan of the University of Missouri, Columbia, who was not involved with the study. When an earthquake strikes, for example, rescue teams could survey the affected area with drones and create a detailed 3D map of the damage in less than a day. The next step for researchers, Palaniappan says, is to improve the algorithm so that it can produce a map within minutes.