Vision Aided Navigation in Unknown Environments by Spencer Ahrens and Dan Levine

Unmanned Aerial Vehicles (UAVs) have demonstrated complex autonomy and aggressive flight outdoors, as well as obstacle avoidance and simple target tracking indoors, but are stringently limited in payload. For surveillance, search & rescue, and reconnaissance missions in urban canyons or indoors, it is then beneficial for the mission payload -- the imaging sensor -- to play a dual role by aiding in localization and mapping.

Example mission environments. Left: a partially collapsed building in Manila, Philippines [1]. Right: remains of NOAA's National Severe Storms Laboratory after a fire [2].

The fusion of six-axis inertial data and low-resolution camera images has enabled rapid prototyping in ACL's RAVEN testbed of novel estimation, planning, and control algorithms for quadrotor vehicles.

Left: an Ascending Technologies Hummingbird quadrotor equipped with a Panasonic BL-C131A netcam in flight.
Right: a new obstacle avoidance algorithm tested in simulation and in flight tests. A repulsion-only trajectory is shown in magenta, and an improved trajectory based on "obstacle corridors" is shown in blue.

These algorithms enable the vehicle to safely navigate cluttered, unknown, static environments without prior information, with limited onboard resources, and in the presence of communication losses.

Left: the hovering vehicle uses image processing to detect features for ego-motion estimation and obstacle avoidance. Right: the vehicle plans around the obstacles in its view and proceeds to fly over the stool.

A complete video of the flight tests can be found here.

Further research will explore planning in the information space of an entirely unknown environment, enhancing the proactive obstacle-avoidance trajectory generator, and autonomously flying through a window or doorway.