Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 119

Fig. 2.13 Probability
distribution image for
camera 1, obtained based on
the histogram of the relative
frequency of greyscale pixel
values

(i) a good three-dimensional pose estimation is available at the first time step and
(ii) that the object appearance does not change considerably over a small period of
time.
To verify an estimated three-dimensional pose, five criteria are used: (i) the tem-
poral variation of the orientation similarity, (ii) the current orientation similarity,
(iii) the temporal variation of the appearance similarity for the three-dimensional
pose T (t) compared with the pose T (t − t) , (iv) the appearance similarity for T (t)

compared with T (t − t) , and (v) the appearance similarity for pose T (t) compared

with pose T (t 2 t) .

If the five criteria of a three-dimensional pose pass a set of thresholds, the pose is
defined as valid and the reference template is added to the database. Otherwise, the
tracking can either be stopped with a warning that the object is lost, or a reference
template-based pose recovery stage can be applied.

−

Pose Recovery on Loss of Object If the object is deemed to be lost by the ver-
ification stage, the pose recovery stage starts at the last valid three-dimensional
pose vector T and computes two possible pose hypotheses: (i) based on a proba-
bility distribution image and (ii) by applying a correlation with the reference tem-
plates.
The first pose hypothesis is obtained using the last 10 reference templates of the
forearm to construct a histogram which describes the forearm appearance. The his-
togram describes the relative frequency of greyscale pixel values (Bradski, 1998 )
and is computed only for camera 1. Similar to Bradski ( 1998 ), the histogram is used
to compute a probability distribution image for camera 1, as shown in Fig. 2.13 .
The last valid three-dimensional pose vector T is used to project a rectangle cir-
cumscribing the forearm into the probability distribution image. A two-dimensional
greedy optimisation in the image plane is applied to find the centre and rotation of
the rectangle with the maximum sum of pixel probabilities. Using the best match-
ing rectangle, we construct a three-dimensional pose under the assumption that the