Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 284

the recognition performance should not decrease in the presence of a cluttered back-
ground displaying arbitrary colours. It is thus favourable to abandon the assumption
of a specific colour distribution separating relevant from background objects.
Many systems for estimating the three-dimensional pose of the human body or
parts of it require a user-dependent configuration and calibration stage. An indus-
trial safety system, however, must be able to detect and localise arbitrary humans
without relying on any kind of a priori knowledge about them, as a large number
of persons may have access to the workspace to be safeguarded by the system. As
a consequence, we will see in the following sections that it is favourable to utilise
person-independent models of the human body and its parts rather than attempting
to adapt a highly accurate person-specific model to the observations.
In the systems described above, multiple hypothesis tracking is a fairly com-
mon technique, but most existing systems merely evaluate two-dimensional cues.
In many cases this is sufficient for interpreting gestures for enabling an interaction
between a human and a robotic system. However, for an industrial system for safe-
guarding large workspaces it is required to acquire accurate three-dimensional data
of the scene of considerably high spatial resolution and assign them to specific parts
of the human body in order to be able to reliably avoid collisions or other hazardous
interferences between the human and the machine.
In the following sections of this chapter, the issue of user-independent three-
dimensional detection and tracking of persons and human body parts not involving
skin colour cues is addressed in the context of safe human-robot interaction in the
industrial production environment.

7.2 Object Detection and Tracking in Three-Dimensional Point
Clouds

This section describes the evaluation of the method for three-dimensional detection
and tracking of persons in three-dimensional point clouds introduced by Schmidt et
al. ( 2007 ) as described in Sect. 2.3 in an industrial production scenario, involving
a workspace with a human worker, a robot, and a moving platform. The presenta-
tion in this section is adopted from Schmidt et al. ( 2007 ). Stereo image sequences
recorded with a Digiclops CCD camera system with an image size of 640 × 480 pix-

els, a pixel size of 7 . 4 m, and a focal length of 4 mm were used. The stereo baseline

corresponds to 100 mm. The average distance to the scene amounts to 5 . 65 m.
A three-dimensional point cloud of the scene is generated by combining the
correlation-based stereo technique of Franke and Joos ( 2000 ) and the spacetime
stereo approach outlined in Sect. 1.5.2.5 as described in Sect. 2.3 . We empirically
found for the correlation matrix element Σ z in ( 2.36 )thevalue Σ z =

μ

0 . 292, regard-
ing a set of three-dimensional points obtained with the spacetime stereo algorithm
and belonging to a plane scene part, while Σ x = Σ y = 1 are equally scaled. The

velocity scaling factor is set to ρ 380 s, where the velocity is expressed in metres

per second and the spatial coordinates in metres. The kernel widths for ( 2.37 )are

=