Structural tracking connects the pixels according to their characteristics and those of their neighbourhood. This method is similar to a grouping such as it is carried out in the processes of region growing, but it integrates characteristics particular to the lineaments, the direction for instance.

From a ``seed'' point, the process analyzes the following pixels and chooses the best candidate to continue the chain. The quality of the result depends on the criteria involved, but also on the preprocessings which were used to initialize the tracking, mainly the detection of the ``seed'' points.

For example:

• S. AIRAULT and O JAMET [S. Airault, 1994], R. RUSKONÐ [Ruskone, 1996] use this technique to extract chains, and optimize criteria of directional homogeneity and distance covered:

1.

From a ``seed'' point, a tree is developed. Each branch represents a straight section of given length and included in a cone of , for which is calculated the longitudinal variance.

2.

Once all the branches developed for a given depth of tree, the optimal path is searched. It must optimize a cost function which tends:

• to minimize the longitudinal variance,
• to maximize the distance covered and
• to maximize the straightness.
  
  

• M.E. of GUNST [de Gunst, 1991] uses a tracking by correlation of profile (introduced by MCKEOWN in 1988). The principle of this tracking is based on the fact that the transverse profile of a road varies very gradually along the road:

1.

From a ``seed'' point and a direction of propagation, the transverse profile of the road is calculated.

2.

Then, at a certain distance from the last point considered and for various directions, included in a cone around the last direction considered, each transverse profile is considered. The propagation follows the direction giving the best correlation of profile.

The major problem of these processes is that they are disturbed by occlusions and often lead to partial chains which it is then necessary to connect.