96 
  
  
ha 
fo 
m; 
wi 
4 
Ti 
ste 
hu 
fas 
itic 
5 
Fig. 7: Segmentation and labeling results with an artificial figure (a) and real images (b, c). The assigned 
limb classes are marked by hatchings in different grey tones: arms — light and long-striped, legs — dark Al 
and long-striped, trunk — cross-striped. Compare with the animated figure where — except for the lower At 
torso part and two arm labels — all possible assignments are made correctly. In the centre of each mark- 
ing a region number is displayed in bright intensity. Ba 
| Bo 
| Ch 
and obey the imposed model constraints. As for the pairwise edge assignment, finding a closed-form solution that 
guarantees consistency and optimality is quite difficult because of the following reasons: | Di 
e The ,filtering effect“ of a previous analysis of primitives’ properties fails to appear because the regions ex- 
tracted are very similar. Either trunk or extremities might be mapped as oblong segments. Yet, extremities and | Do 
trunk may be fractured into short parts like the reconstruction of the lower trunk of the synthetic figure in | Ed 
fig. 7a. | Fle 
€ Ifa relaxation is beginning with all regions bearing all possible labels, in the end a unique assignment can sel- | H 
dom be achieved with the data at hand. A subsequent optimization step would have to solve competing solu- | a 
tions. | Hi 
Thus, the algorithm chosen only roughly follows a general labeling scheme. An upright standing or moving person | Ho 
is assumed for the moment: | 
1. First, a region cluster is divided horizontally into two | Kir 
parts if more than five regions are present. Ideally, this m | 
. would separate leg regions from all others. Lei 
2. Each set of regions is analysed with regard to their widths, 
geometric connections at front sides and similarities in Ok 
angle and intensity. Then, neighbouring elements are | 
grouped into units and potential class values are deter- | Ox 
mined (without performing a final assignment at that | 
stage). | Per 
3. Inthe end, the collected evidence for each region is com- 
bined to form an object hypothesis. It turns out that the 
variance in width deviation between potential leg regions | Qui 
is smallest. A topology check guarantees consistency with | Rol 
the object model. | 
Having assigned a label — three object classes, c.f. fig. 7, and | 
a rest class — to each region, the data-driven image analysis is O'R 
completed. 
Additional tests, now under model control limiting the | Seg 
search space, are performed to further validate the object hy- Fig.8: Round contour lines found by pixel | 
pothesis. For example, it is checked with circular masks as to ~~ comparisons around circular search paths. 
whether the mainly round head or shoulder areas are present Positive evidence is marked by squares of 
in specific image locations. Fig. 8 shows where round contours 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995 |