Applications to Safe Human–Robot Interaction - 3D Computer Vision: Efficient Methods and Applications - page 294

7.4.4 System Configurations Regarded for Evaluation

Several system configurations are evaluated by Hahn et al. ( 2010a ), including dif-
ferent combinations of three-dimensional pose estimation and tracking modules.

Configuration 1: Tracking Based on the MOCCD The first system configura-
tion is based on the MOCCD algorithm. To start tracking, a coarse initialisation of
the model parameters at the first time step is required. We apply three instances
of the MOCCD algorithm in a multi-hypothesis Kalman filter framework. Each
MOCCD instance is associated with a Kalman filter, where each Kalman filter im-
plements a different kinematic model assuming a different object motion. The idea
behind this kinematic modelling is to provide a sufficient amount of flexibility for
changing hand-forearm motion. It is required for correctly tracking reversing mo-
tion, e.g. that which occurs during tightening of a screw. A winner-takes-all com-
ponent selects the best-fitting model at each time step using the following criteria:
(i) the confirmation measurement, (ii) the quality of the prediction, and (iii) the
difference of the probability distributions of the pixel grey values along the model
curve. The confirmation measurement is introduced by Hanek ( 2004 ) and is an in-
dicator of the convergence of the optimisation. The second criterion describes the
similarity of the prediction and the pose measurement. With the third criterion it
is ensured that the MOCCD algorithm separates probability distributions along the
projected curve. A hypothesis that is better than any other in at least two criteria is
deemed the winner. It is important to note that the Kalman filter is only used for
initialisation purposes, while the evaluation refers to the actual MOCCD measure-
ments.

Configuration 2: Tracking Based on the Shape Flow Method The second con-
figuration is based on the shape flow (SF) method. Similar to configuration 1, track-
ing starts with a user-defined parameter vector T (t

1 ) . A hierarchical approach
is used to determine the three-dimensional pose T and its temporal derivative T .
The three-dimensional pose T (t) at time step t is computed using the MOCCD
algorithm. The temporal pose derivative T (t) at time step t is determined with the
SF algorithm as described in Sect. 2.2.3.4 using the images at the time steps (t

=

t)
and (t − t) . To achieve a temporally stable and robust tracking, we rely on two
pose hypotheses, where a winner-takes-all component selects the best-fitting spatio-
temporal model at each time step using the same criteria as for configuration 1. The
prediction of the parameter vector for the first hypothesis is computed with a con-
stant velocity model, and for the second hypothesis we apply a constant position
model. These predictions are used at time step (t

+

t) as initialisations for the
MOCCD and SF instances. The idea behind this two-hypothesis approach is to pro-
vide a sufficient amount of flexibility for changing hand-forearm motion without
requiring a large number of MOCCD and SF evaluations.

+

Configuration 3: ICP-Based Tracking This configuration applies an ICP-based
'tracking by detection' approach for which no pose initialisation at the first time