Comtuter and Internet News Channel
Autonomous navigation system based on vision will improve the mobility of military drones and civilian drones-”postmen”.
Central scientific-research and experimental design Institute of robotics and technical Cybernetics (RTC) at the request of the defense Ministry of Russia is developing a new orientation system for aerial, terrestrial and underwater drones. As report “News”, test equipped its drones has already begun.
Currently, the orientation of the drone in space, it uses either the GPS (satellites), or inertial system gyros, “consider” their speed depending on the movement of the device in one direction or another. Satellites are unreliable because in bad weather, their signals may “not pass”, but under the water or in close development the use of them may not be under the clear sky. Inertial navigation without external correction can quickly accumulate errors, and is really a fine device for her rather bulky and expensive.
The best solution would be to create a third, correcting navigation system that is not dependent on external factors — for example, technically reproducing how focused on the man himself. According to the developers, their computer vision system is comparable in its capabilities with the human eye. She will be able to compare photographs of the area from popular online services (Google and Yandex) that “sees” itself using stereo cameras, which will give the opportunity to effectively navigate the terrain.
At the moment, experiments on Autonomous navigation for military drones has already begun. It is planned that the first robots with a new, Autonomous control system will go into force by 2021 (first in pilot operation). Among the main obstacles on this way is a refinement of the vision systems to a comparable level with the person and Supplement it with elements of artificial intelligence. It is indispensable in the “identification” area on the photos.
The project created an algorithm that will allow drones to operate in the group. In the area of task execution, a single UAV will navigate through the existing photo, draw their own pictures of the area and to apply the new data with the new landmarks to the map, are available by combining into a single network for all devices. With each new flight on the same area of his three-dimensional map will become more detailed. In theory, the reference systems for the reference points in flight will provide enough accuracy for navigation of UAVs. At least, if the map data for this area is quite urgent.
Similar developments are under way also abroad. In the last decade, improving methods of so-called concurrent mapping and orientation data from one or more cameras (SLAM — from the English. Simultaneous Localization and Mapping). One of the main obstacles on this path is that vision systems usually do not have enough “intelligence”. The two most well-known incident in which it appeared — the death of two drivers of the Tesla. The first happened in 2016 due to the fact that the autopilot in the first case, could not distinguish white tank truck from the bright daytime sky, against which he saw the tank. Second — because of the “changing landscape” — in the course of the previous accident the truck broke the concrete separation on the track, and the car failed to take into account this change.
Therefore, to create a truly effective vision we need a system of “artificial intelligence”, able to take into account the situation described above. Need some non-trivial solutions. The same Tesla actively uses neural networks, trained based on the dataset obtained from the drivers-the people (through the camera loading video driving everyday drivers on the server Tesla). However, for Autonomous navigation of UAVs all the more complicated. Often they are unable to benefit from the experience of human operators because the environment in which they operate, much more difficult traffic situations, and the mathematical formalization of what is happening to flying a drone is much more complicated.
While Autonomous navigation is developed in our country, only for military drones, there is no doubt that a similar problem will soon become relevant for civil drones. The recent crash of the UAV “Mail of Russia” indicates the need to improve the technical vision for a “peaceful” drones — for example, delivering small goods and parcels in urban areas, where dense development often makes the orientation of only one of the satellites is unreliable.