The basis for the use of unmanned aerial vehiclesobjects laid technology that allows the device to detect and avoid obstacles in time. However, the speed of maneuver does not always satisfy users. Swiss engineers have developed navigation devices that allow the drone to perform rapid maneuvers to avoid a collision with a ball thrown at it.
At the present stage of development of navigationdrones, it takes 20 to 40 ms to start responding to an emerging obstacle. Such parameters are permissible when maneuvering around a stationary or slowly moving object. However, when approaching another drone or bird, the device does not have time to work out the evasion, and a collision becomes inevitable.
Thus the slow reaction of dronesgives rise to many problems when it is necessary to use a group of drones, for example, during a survey of the area affected by a natural disaster. Rescuers to conduct effective measures to rescue survivors require a quick survey of a large area using several drones, which may collide with a weak reaction of the navigation system to approaching high-speed objects.
Developers from the University of Zurichattempted to solve this problem. The usual serial quadrocopter was equipped with video cameras recording fast movements and working according to special algorithms. Through the efforts of Swiss engineers, the preparatory phase for maneuver was reduced to 3.5 ms.
The new system uses specialevent cameras - sensors of dynamic vision. Traditional CMOS cameras use pixel-based tracking of the entire scene. In this case, the algorithm tracks each pixel and its change with respect to neighboring pixels. The event camera algorithm individual pixels are turned on and off independently when a change in light intensity is fixed to a threshold value. Thus, the main part of the fixed pixels is in a "sleeping" state and does not load the software and processing power of the system. This all leads to an increase in the speed of processing the information received from the cameras.
To date, serial drones have notused event cameras and their software does not support working with such sensors. As a result, engineers from Zurich first of all created a software for event cameras that works in real time and allows you to track changes in the surrounding space and respond in time to an emerging obstacle.
At the first stage of testing, developersindependently threw various objects in the direction of the event camera in order to find out how efficiently the algorithm is able to recognize objects. As a result, the recognition efficiency of an approaching object was from 81 to 97%, depending on the distance to the object and its size.
At the next stage, the developers investigatedthe reaction of two cameras mounted on a drone that flew both in the pavilion and outside. Two cameras are needed to realize the stereoscopic vision of the drone.
The experiment showed that the drone equippedevent cameras and working on the algorithm of Swiss engineers, is able in 90% of cases to avoid a collision with a ball launched from a distance of 3 meters at a speed of 10 m / s.
The introduction of such technology on serial droneswill increase the speed of devices by almost 10 times, without increasing the potential danger of a collision. This will help to increase the efficiency of using unmanned aerial vehicles during rescue operations, package delivery and other tasks.