Stereo Detection System Lets Autonomous Drones Fly Faster

Algorithm brings high-frame-rate processing to UAVs.

Unmanned aerial vehicles (UAVs) have the potential to automate much of the working world, from search and rescue to home deliveries and building surveys.

Drone autonomously avoiding obstacles at 30 MPH. (Video courtesy of CSAIL)

However, before they can work for us, UAVs need to be able to negotiate our obstacle-filled world by themselves without incident.

To that aim, researchers at MIT have developed a new algorithm called “pushbroom stereo” that enables autonomous drones to navigate cluttered environments at relatively high speeds.

Algorithms for obstacle detection normally search for objects at multiple depths. This makes them computationally intensive, which requires heavy onboard computing power. This plus the weight of sensing equipment explains why autonomous drones have historically been limited to speeds of five to six miles per hour (eight to nine kilometers per hour).

In-flight snapshot of single-disparity stereo detections on a goal post (blue boxes) and past detections (red dots). Airspeed is shown in MPH (left) and altitude in feet (right). (Image courtesy of CSAIL)

In-flight snapshot of single-disparity stereo detections on a goal post (blue boxes) and past detections (red dots). Airspeed is shown in MPH (left) and altitude in feet (right). (Image courtesy of CSAIL)

Pushbroom stereo reduces the computational complexity of obstacle detection by searching for objects only at a depth of 32 feet (10 meters). This information is then integrated with data from the odometer to “push” earlier detections forward as the drone flies, like a broom pushing debris. That lets the algorithm quickly build a full 3D map of obstacles in front of the drone. As a result, the drone can safely travel at speeds of up to 30 miles per hour.


You might think that searching for objects at only a single depth would entail a significant loss of information. The algorithm’s creator, Andrew J. Barry, argues that this is actually an advantage. “You don’t have to know about anything that’s closer or further than that. As you fly, you push that 10-meter [32 feet] horizon forward, and, as long as your first 10 meters are clear, you can build a full map of the world around you,” he said.

The researchers claim that their drone employs the first high-frame-rate stereo-detection system running onboard a small UAV. However, they admit that their approach results in occasional incorrect estimates or “drifts.” This issue will be resolved as lightweight computational power increases.

In order to test the algorithm, Barry and his supervisor, Russ Tedrake, implemented it on a quad-core 1.7GHz ARM processor and two eight-bit grayscale cameras running 376 x 240 pixels at 120 frames per second. The hardware was mounted on a delta wing drone with a 34-inch (86-centimeter) wingspan.

The total cost of the test drone was $1,700 using off-the-shelf components. This is nearly half the cost of autonomous drones that are commercially available today.

For more information, visit MIT’s CSAIL website.

Written by

Ian Wright

Ian is a senior editor at engineering.com, covering additive manufacturing and 3D printing, artificial intelligence, and advanced manufacturing. Ian holds bachelors and masters degrees in philosophy from McMaster University and spent six years pursuing a doctoral degree at York University before withdrawing in good standing.