A drone can lose satellite navigation in mountains, "gray zones," when the signal is jammed, or in radio silence mode. In such conditions, it effectively loses its usual reference point. MIPT has solved the problem and created the "Fasetka" system, which helps the device understand where it is, even without GPS and GLONASS: it looks at the terrain, recognizes stable objects, and cross-references them with a pre-loaded map.
The development was created at the MIPT Robotics and Computer Technologies Research Center with the support of the NTI Foundation. The system is designed for use where conventional satellite navigation is unavailable or unreliable: in disaster areas, technological accidents, for pipeline monitoring, and for special tasks in radio silence conditions.
Tests on a helicopter and drones confirmed that at an altitude of 1000 meters, navigation accuracy can reach 6 meters.
Conventional visual orientation methods also have limitations. A drone can use pre-prepared maps, but the landscape changes: in winter, snow covers familiar contours, at night and in bad weather, many systems work worse, and angled shooting greatly complicates recognition. "Fasetka" solves these problems.
The device collects different types of terrain data and searches for objects that change little over time: roads, buildings, riverbeds, power line poles. For example, the thermal contour of a building can be preserved in winter, and a radar image helps to see the large structure of a forest regardless of foliage. These features are then compared with a reference satellite map loaded into memory – for example, Google or Yandex Maps.
Another important feature is operation at an angle. Classical visual navigation usually requires the camera to look strictly down. "Fasetka" allows deviation up to 45 degrees from the vertical. This was achieved by training the neural network on synthesized angles and using generative models.