Detail Search : Model Matching

Introduction

Knowing the axis system fitted to the vehicle given by first part of the processing schema, we know where given details (stored in a database) should appear on the image. This position is approximate because :
• The axis system is known to a limited precision.
• The detail shape is known to a limited precision.
A typical example is a wheel stored in the database with a (reasonable) error on the radius.
• ... .
The initial position is thus used as an initial guess and the goal of the model matching is to search in the 'neighbourhood' (position, scale, ...) if the detail is present.

Search Space

We said we search the detail in the 'neighbourhood' of the initial position. We have to precise what that neighbourhood is, we have to define the search space.

The error sources described in the introduction implies the we have to search for the position but also for some shape transformation.

The allowed transformations of the detail are all linear axis transformations which include translation, rotation, scaling and sharing.

Search Strategy

Experience has shown that it is hopeless to search a detail if neither the position nor the shape are known with reasonable precision. We hypothesis that the shape (scale and sharing) of the detail are known to a reasonable accuracy and the detail is first located without modification of shape.

Detail Localisation

Different search methods are used according to the detail type:
• Big details (relatively to the initial guess error) as the track, the wheels,... have an extended attraction basin (the detail should be 'felt' from the original position) and an Iterated Conditional Mode (ICM) optimisation is used.

• Small details as the support wheels have a reduced attraction basin and the detail is first located by means of a hough transform which return a number of candidates. The best one will be selected as the one giving (after shape transformation) the best energy.

Detail Shape Transformation

After localisation, the shape of the detail may be searched

• by searching for the best scaling and sharing by means of an ICM optimisation.

The problem is to define appropriate centre and directions for those transformations. This is feasible for shapes presenting preferential directions as the tracks and has shown high efficiency in practice.

• by searching for the best linear transform

For other shapes as the wheels, the 'best transformation' (in the least mean square sense) is computed and applied.

Results

Visual Evaluation

On the following image, an example of initial conditions (red) and found (blue/green) detail (the axis system was intentionally perturbed: sigma=10).



Numerical Evaluation

On the following charts, search results are shown for increasing axis noise. Fore each axis noise, 10 samples are shown.