IEEE Robotics & Automation Magazine - December 2013 - 23

A camera is a lightweight sensor with low power con- algorithm [16] is adopted for outlier rejection [17]. A new stasumption. Different vision-based obstacle-avoidance tech- tistically robust triangulation-based approach is then proposed
niques are implemented in this article, most of which are to estimate the corresponding 3-D points, which are used to
based on the biologically inspired concept of optical flow. In build a 3-D occupancy grid map of the obstacles in the sur[4], fruit flies avoid obstacles when turning away from a rounding environment.
region with high optical flow, while in [5], honeybees flying
The map is updated by adding the data provided by a target
through a tunnel try to balance the amount of lateral optical identification algorithm, based on an elaboration flow similar
flow to keep the same distance from the walls. By exploiting to the obstacle localization method. Based on this map, a systhis concept, a fisheye camera pointing downwards is tem of repulsive forces from obstacles and attractive forces to
employed in [6]-[8] to get image frames and to compute the the target is built. This force is then applied to generate a feedlateral optical flow: by using this information, a depth map of back force on the haptic device, provide the human operator
a corridor is built to control the UAV along a trajectory cen- with a perception of the environment, and generate an additered in the middle of the observed obstacles. In [9], two opti- tional position reference for the UAV controller. With this new
cal flow controllers for a fully actuated UAV that flies in con- navigation support system, the human operator is able to steer
fined indoor and outdoor environments are used: the first the UAV to the identified target while avoiding obstacles along
controller regulates the UAV's forward thrust, and the second its trajectory. In the experimental validation, the UAV pose
one controls the UAV's side-slip thrust.
(position and orientation) is stabilized by means of an external
Two different approaches with and without optical flow are motion capture system to achieve higher robustness.
proposed in [10], and are based on the concept of the time-tocontact needed before the vehicle collides with an obstacle System Overview
while the UAV is moving with a specific translational velocity. In this section, the main functionalities of the developed teleA triangulation-based method for the computation of the dis- operation system are described, as depicted in Figure 1.
tances of detected objects is shown in [11], where a stereo
camera system is used. Although stereo images benefit in The Haptic Interface
terms of accuracy of the estimation and odometry of the The human operator interacts with a haptic device to give a
obstacles position, a high payload and computational power is reference command to the UAV. At the same time, the operarequired from the UAV. On the other hand, the main draw- tor can feel a force feedback, which provides a haptic feeling
back in the use of a monocamera system is the need of a trans- of the identified environment, i.e., how close the UAV is to the
lational movement of the UAV to detect the surrounding localized obstacles and target.
obstacles. In [12], a single-camera front-obstacles collisionavoidance approach for a microflyer is presented based on the The Ground Station
divergence of the optical flow on the left and right side of the In the proposed framework, the ground station has to: 1)
direction of travel.
send the desired reference to the onboard controller, 2) generIn this article, two main problems are addressed: 1) a ate the force feedback to be displayed by the haptic device, 3)
monocamera vision-based obstacle avoidance for a safe navi- estimate the UAV's odometry by using the images acquired by
gation of a bilateral teleoperated UAV by means of a haptic the onboard camera, and 4) process the images to build a
device that endows the human operator with a vision-based local map of the environment. Based on the map and the estihaptic feedback of the environment and 2) an image target mated state of the UAV, a virtual interaction force applied to
identification system to generate a force field that leads the the UAV is built when obstacles and/or the target are
human operator to safely reach the target. This new approach detected. This force arises when either an obstacle enters in a
makes use of a single onboard calibrated camera together with a
monocamera odometry estimation
UAV
system based on Parallel Tracking
Communication
and Mapping (PTAM) [13]. The
Channel
Ground Station
images, grabbed by the onboard
Haptic Device
camera, are stored in a buffer
Camera
together with the corresponding
UAV position estimation, i.e., the
UAV Pose
UAV odometry. Then, robust
Estimation
image keypoints are extracted and
matched between image pairs by
means of a Pyramidal LukasHuman Operator
Kanade algorithm [14], [15].
Moreover, a Random Sample
Consensus (RANSAC)-based Figure 1. A block diagram of the overall teleoperated control system.
DECEMBER 2013

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2013