As can be seen in Figure 19, even a judiciously tuned edge segment extraction process provides a multitude of DESs which may complicate the search for appropriate associations between DESs and MESs. But quite apart from this problem, another one may turn out to be even more difficult, namely the fact that the polyhedral object model does not contain any information about the direction of the expected gray value gradient across an MES. As a consequence, the Mahalanobis-distance between a DES and a correctly associated MES may turn out to be larger than the Mahalanobis-distance between such a DES and a tentatively, but incorrectly associated MES. Wrong associations between DESs and MESs can influence the state vector update in such an infortunate manner that the update process - or subsequent steps in the Kalman-Filter based tracking process - may diverge.

**Figure:** Data and model edge segments computed for a region from a frame of the sequence depicted by
Fig. 25. The long model segment
from the contour of the bus roof (denoted by `Modellkantensegment')
with a length of 104.7 pixels has been associated to the short
data segment (denoted by `Datenkantensegment') with a length of
only 5.3 pixels. See text for a detailed explanation of this phenomenon..

As a consequence, the two error-prone subprocesses of heuristic aggregation of EEs into DESs and the combinatorial search for appropriate associations between DESs and MESs has been abandoned. The SS is turned into a `synthetic image gradient' by convolving it with a 2D-Gaussian. Figure 22 illustrates the result. This `expected synthetic gradient' image is compared to the actually observed edge element image as shown in Figure 18 in order to determine which modification of the state vector may reduce the sum of the squared pixelwise difference between the image gradient magnitude and the synthetic image gradient magnitude [Kollnig & Nagel 95].

**Figure:** Synthetic Gradient Norm for the SS shown in Figure 12.

As can be seen in Figure 23, the update loop becomes conceptually simpler - albeit at the price of increased computations. The real gain, however, consists in the increased reliability of this approach as compared to the one illustrated by Figure 4.

**Figure:** Interaction of the detection - initialisation - actualisation process using a synthetic gradient norm instead of edge segments as in Figure 4 .