In the last decade researchers have built incredible new capabilities for small aircraft, with quadrotors moving from labs to toy stores and with autonomy reaching smaller and smaller vehicles. As the systems, and their payload capacities shrink, we can no longer use typical aircraft sensors such as RADAR, scanning LIDAR, and other active sensing methods for obstacle detection and avoidance. Smaller vehicles must move to lighter weight sensors. Cameras are lightweight, fast, and information dense but can require sophisticated, often slow, processing to be useful for robotic applications. Here, we demonstrate that highspeed embedded stereo vision offers a way forward for fastflying aerial vehicles. The presented solutions are lighter, provide more information, and use less power than laser ranging systems. They work in outdoor environments that overwhelm active IR depth cameras, and do not suffer from scale observability like monocular camera systems. The new algorithms and processing techniques presented here enable us to detect obstacles faster, with less payload, and less computational hardware than ever before. We present the only two embedded stereo (two-camera) vision systems running at high frame rate with low enough power and weight requirements for small flying platforms. These systems provide full 3D positions of points seen by the cameras, which is essential for obstacle avoidance and safe, robust flight. These two systems were developed independently by ETH Zürich and MIT for solving the problem of providing accurate, rich sensing in a package small enough to fit in the payload constraints of mico-aerial vehicles (MAVs). The two systems differ in their processing, with one using an onboard FPGA (field programmable gate array) to perform dense semi-global matching (SGM) [4], [6] and the second using conventional ARM processors to perform pushbroom stereo [1]. Both systems run at 120 frames per second (fps) at 320x240 pixel resolution, on duplicate fixedwing platforms flying over 30 MPH (13.4 m/s).
[1]
Roland Siegwart,et al.
A synchronized visual-inertial sensor system with FPGA pre-processing for accurate real-time SLAM
,
2014,
ICRA 2014.
[2]
Marc Pollefeys,et al.
Real-time and low latency embedded computer vision hardware based on a combination of FPGA and mobile CPU
,
2014,
2014 IEEE/RSJ International Conference on Intelligent Robots and Systems.
[3]
Stefan K. Gehrig,et al.
A Real-Time Low-Power Stereo Vision Engine Using Semi-Global Matching
,
2009,
ICVS.
[4]
Markus H. Gross,et al.
An FPGA-based processing pipeline for high-definition stereo video
,
2011,
EURASIP J. Image Video Process..
[5]
Marc Pollefeys,et al.
Reactive avoidance using embedded stereo vision for MAV flight
,
2015,
2015 IEEE International Conference on Robotics and Automation (ICRA).
[6]
Russ Tedrake,et al.
Pushbroom stereo for high-speed navigation in cluttered environments
,
2014,
2015 IEEE International Conference on Robotics and Automation (ICRA).
[7]
Marc Pollefeys,et al.
An open source and open hardware embedded metric optical flow CMOS camera for indoor and outdoor applications
,
2013,
2013 IEEE International Conference on Robotics and Automation.
[8]
Dario Floreano,et al.
Near-obstacle flight with small UAVs
,
2008
.
[9]
Toby P. Breckon,et al.
An empirical comparison of real-time dense stereo approaches for use in the automotive environment
,
2012,
EURASIP J. Image Video Process..