A team of researchers led by Prof. Davide Scaramuzza have developed a way to train drones to follow forest trails in an effort to assist search and rescue missions for lost hikers. According to the research, Prof. Scaramuzza's team figured out a method of machine learning through Deep Neural Networks (DNNs) which enables an unsupervised drone to determine the direction of a path using an on-board camera.
The system was created by first setting up a hiker with three cameras that cover about 180 degrees of visual information: one positioned straight ahead, one placed 30 degrees to the left and the other 30 degrees to the right so that there is a slight overlap in the captured video. The hiker was instructed to always look ahead in the direction of the path since the front camera will provide the information for the trail.
The raw data (PDF) used was eight hours' worth of footage of approximately 7 kilometers of hiking trail between an altitude of 300 and 1,200 meters. The footages were taken at different times of the day and under different weather conditions.
The results were surprising when it was tested because the autonomous quadcopter was able to navigate a completely new trail and stay on course as well as, and sometimes even better, than humans. The same path and test was done with two humans against the drone to determine how effective the DNNs based machine learning was and, on one test, the quadcopter was successful 85.2 percent of the time as opposed to the two people who were accurate 86.5 and 82 percent of the time. A second test with different conditions resulted in the quadcopter being accurate 95 percent of the time when the two people were 91 and 88 percent accurate.
Watch the video explanation of the research below.
"Now that our drones have learned to recognize and follow forest trails, we must teach them to recognize humans," Prof. Scaramuzza said.
A drone that could recognize proper trails and humans will certainly be of great assistance to rescue operations, moreso if it can also detect vital signs like the Lynx 6-A.
The research titled "A Machine Learning Approach to Visual Perception of Forest Trails for Mobile Robots" appeared in the IEEE Robotics and Automation Letters (RA-L) and will be presented during the IEEE International Conference on Robotics and Automation (ICRA'16) in May.
