Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 277

tures performed by humans in the application context of human-robot interaction,
concentrating on pointing gestures and the interaction with objects.
7.1.1 Vision-Based Safe Human-Robot Interaction
In the safety system described by Vischer ( 1992 ) in the context of a robotic sys-
tem for object manipulation, the expected position of an industrial robot is verified,
and persons in the neighbourhood of the robot are localised in a camera image.
A background image is subtracted from the current image of the workspace, where
a method which behaves robustly with respect to changing illumination is applied.
A robotic system for grasping objects is described by Schweitzer ( 1993 ), which
partially builds upon the work by Vischer ( 1992 ). Schweitzer ( 1993 ) points out that
this system is equipped with two safety systems developed by Baerveldt ( 1992 ),
where one safety system consists of accelerometers which detect exceedingly fast
robot movements, while the other system is based on a camera, comparing the ex-
pected robot position to the position inferred from the image and estimating the
position of a human in the neighbourhood of the robot. According to Schweitzer
( 1993 ), the robot may be stopped if necessary, or an alternate path can be defined in
order to avoid endangerment of the human.
Ebert and Henrich ( 2003 ) describe the SIMERO system for safe human-robot in-
teraction, which provides an integrated image-based framework for safeguarding the
collaboration between a human and an industrial robot. A background subtraction
is performed for the acquired images of the workspace based on a reference image,
resulting in a silhouette of the human. A robot modelling stage yields a silhouette of
the robot relying on its joint angles. The extracted silhouettes of the human and the
robot are used to predict collisions without extracting the three-dimensional struc-
ture of the scene. Subsequently, the path of the robot is planned.
Gecks and Henrich ( 2005 ) regard the problem of 'pick-and-place operations'
during which objects are manipulated, leading to changes in the background or ref-
erence image used for object detection. Hence, an algorithm for adapting the refer-
ence image to changed object positions is introduced. Based on the object detection
results, a robot path is determined that does not lead to collisions. The system relies
on several cameras, but does not perform an actual three-dimensional reconstruction
of the workspace.
The safety system of Kuhn et al. ( 2006 ) concentrates on movements of a robot
guided by a person, where several camera images are combined with data of a
force/torque sensor attached to the robot. Object detection is performed based on
background subtraction of each image, and the three-dimensional positions of the
objects are estimated. A collision-free robot path is determined, where the maxi-
mum permissible robot speed depends on the relative positions of the robot and the
objects.
Henrich et al. ( 2008 ) propose a safety system which relies on several cali-
brated cameras used to obtain a three-dimensional reconstruction of the workspace.