Distress and rescue drone capable of autonomous flight without hesitation in the forest with deep learning

In the development of robots that can autonomously fly in forests and mountains, traditional research focused on techniques that can distinguish the road itself. Analysis of the images of three action cameras at once by analyzing images of three action cameras at the same time assumes that sorting of routes based on such image features and appearance contrast is inappropriate as autonomous running technology of robots, and the accuracy of the correct route comparable to humans New technologies that can be selected with.

On the Visual Perception of Forest Trails
http://people.idsia.ch/~giusti/forest/web/

Switzerland's Artificial Intelligence Laboratory, which is researching technologies capable of identifying roads such as mountains and forests by analyzing camera information visual information by deep learning,IDSIA"And a collaborative research group by the Robotics Perception Research Group of the University of Zurich. It is easy to understand what kind of mechanism the forest rescue new technology that can be installed in the drone is together in the following movie.

Quadcopter Navigation in the Forest using Deep Neural Networks - YouTube


The new technology installed in Quad Cotter is "to automatically fly by following the way in the forest" "Can be active in the search and rescue mission of outdoor victims" "Identify the way with a commercially available color camera (GoPro) It has features such as.


The way in nature is hard to distinguish even by humans, and it is not possible to realize robots from the images and images with the correct accuracy with conventional technology. The biggest aim is to establish a technology that allows you to select the right direction in places where you can see street like left, front and right as follows.


To solve the problem, the collaborative research group of IDSIA and the University of Zurich analyzed images of three cameras by deep learning and automatically determines the correct course.


When the drones are flying over the road, the system captures images from the camera in 10 layersneural network, 150 thousand loads, 500 thousand neurons, 57 million connections.


In order to make the system installed in the drone discriminate the right way, three GoPros are attached to the human body in advance and learning data is acquired.


The number of pictures entered as learning data has reached 20,000.


By processing these data, we can correct the course appropriately in the three directions visible to the camera, so we can autonomously take the right course without hesitation.


As a result, the course identification accuracy by the new technology is 85.23% compared with the conventional 52.32%, which means that it achieves a discrimination accuracy of 82% which is comparable to that of humans.