As drones become more widespread, the ability to track and monitor them becomes more important. I've written previously about various projects that aim to provide such "air traffic control," such as the automated approach proposed by the Stanford Intelligent Systems Laboratory last year. The team believes that an automated approach to air traffic control is vital because the sheer number of drones in the air make replicating the human-based approach is simply not scalable.
Of course, one of the main challenges of such scalability is the ability to keep an eye on such large numbers of craft. A team from the Ecole Polytechnique Federale de Lausanne (EPFL) has recently demonstrated how a straightforward camera can be used to detect and track drones.
This year has seen a growing number of near-misses between drones and other air traffic. The drones themselves usually lack any means to locate other traffic, so the emphasis has to rest elsewhere to ensure accidents don't occur. The EPFL team developed a range of algorithms to detect and track drones using nothing but an off-the-shelf camera.
The team is working to turn their proof of concept into a real-time detection and collision avoidance system using funding from the Commission for Technology and Innovation (CTI).
Such an approach is necessary because the current method of collision avoidance requires crafts to actively calculate their positions and communicate this information to other craft who can act appropriately. It only works if all crafts have the same equipment, which when talking about drones is simply not the case.
Cameras can help to bridge that gap, but the team needed to create a system whereby the camera can spot another moving object. On a drone, it's especially challenging because of the range of motion a drone has, plus the speed with which detection needs to occur, with this often requiring detection to occur when drones (which are rarely the same shape) are far off in less than ideal visibility. In other words, it's not easy.
The work, which was documented in a recent paper, shows that this can be a surmountable challenge, however, with AI helping the camera learn what is, and is not, a drone. This process required a significantly larger database of images with which to train the system. To do this, the team generated realistic synthetic images were merged with the real images to train the system.
Using this approach, the team were able to create a reliable detector using the same kind of lightweight camera found in most smartphones today. The hope is that with further training using an even larger database, the performance will improve both in terms of speed and accuracy. The team hopes that the first commercial models will be released in a years time.
It's an interesting project, and you can learn more about it via the video below.