Algorithm helps drone avoid crashes

A US-based student has developed an algorithm that helps drones fly through trees at 30 miles per hour without crashing.

WESTERN MASSACHUSETTS, UNITED STATES (MIT COMPUTER SCIENCE AND ARTIFICIAL INTELLIGENCE LAB) – An algorithm generated obstacle-detection system that allows a drone to autonomously dip, dart and dive through trees at 30 miles per hour, without crashing, has been developed by a US researcher.

Andrew Barry developed the system as part of his PhD thesis at the Massachusetts Institute of Technology’s (MIT) Computer Science and Artificial Intelligence Lab (CSAIL).

His algorithm allows for real-time object-detection without the need for remote sensing technology, enabling a full map of the drone’s surroundings to be built in real-time. The open-sourced software operates at 120 frames per second, extracting depth information at 8.3 milliseconds per frame.

Other algorithms that have attempted to avoid obstacles are computationally intensive, meaning drones cannot safely fly faster than around five miles per hour without specialised processing hardware.

Another difficulty is that cameras possessing the necessary frame rate and resolution for use in UAVs require an inordinate number of pixels, whereas Barry posited that a drone camera doesn’t need to be concerned with most of what it sees. Therefore, only a small subset of measurements, i.e. distances of 10 metres away, are required for processing.

“You don’t have to know about anything that’s closer or further than that,” said Barry. “As you fly, you push that 10-metre horizon forward, and, as long as your first 10 metres are clear, you can build a full map of the world around you.”

According to Barry, the software can quickly recover any missing depth information by integrating results from the drone’s motion sensor data and previous distances flown, allowing flight to continue without the risk of collision.

Barry hopes to further improve his algorithms so that they can work in thick forests and other challenging environments.

He said: “As hardware advances allow for more complex computation, we will be able to search at multiple depths and therefore check and correct our estimates. This lets us make our algorithms more aggressive, even in environments with larger numbers of obstacles.”

The drone weighs little more than a pound (0.45 kilograms) and has a 34-inch wingspan. Made from off-the-shelf components including a camera on each wing and two processors it cost about $1,700 USD to make.