Drones learn autonomic navigation from cars and bicycles
For navigation, commercial drones use GPS, which works well at high altitude. But what happens when the drones fly independently between buildings or in the dense road network, where cyclists and pedestrians can suddenly cross their way? Until now, commercial drones have not been able to react quickly to such unforeseen events.
Researchers from the University of Zurich and the Swiss National Research Competence Center NCCR Robotics have now developed the DroNet algorithm, which can safely guide drones through the streets of a city. This was set up as a fast residual network with eight levels and generates two outputs for each input image: one for navigation to fly around obstacles and one for the probability of collision to detect and react to dangerous situations. DroNet detects static and dynamic obstacles and reduces speed to prevent collisions. With the algorithm, the researchers see themselves one step closer to the goal of integrating independently navigating drones into our everyday life.
The Swiss researchers’ drone uses a normal camera such as that of a smartphone, and a powerful algorithm for artificial intelligence to evaluate the observed situations. This algorithm consists of a so-called Deep Neural Network. The software learns to solve complex tasks using numerous training examples. “It shows the drone how it solves certain tasks and difficult situations,” explains Davide Scaramuzza, professor for robotics and perception at the UZH.
One of the biggest challenges for Deep Learning is to collect several thousand of such training examples. In order to gather enough data, the researchers collected travels of cars and bicycles that navigated in urban environments and complied with traffic regulations. By imitating, the drone learned to follow rules such as “How to follow the road without getting into oncoming traffic” or “How to stop when obstacles, such as pedestrians, construction sites or other vehicles block my way”. The researchers were also able to show that their drone was not only able to navigate through roads, but also found its way into completely different environments for which it was never trained – for example in multi-storey car parks or office corridors.
The study shows the potential of drone deployments for surveillance tasks or parcel deliveries in busy environments as well as for rescue operations in the event of urban catastrophes. However, the research team warns of exaggerated expectations of what light, low-cost drones can do. “Many technological problems still have to be solved before the most ambitious applications can become reality,” explains PhD student Antonio Loquercio from the UZH.
Original Paper and video, published by the UZH: https://rpg.ifi.uzh.ch/dronet.html
Related articles:
Multi-task deep learning algorithm speeds up in-video analytics
videantis claims 1000x performance boost on deep learning applications
EPFL Drone Days: a showcase of flying technology and robotics
Applying V2X technology to drones
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
