The mines may be hidden in any terrain. The landscape ranges from a
sandy beach to a rocky hill with bushes, trees, ..., etc. The
ground composition may be anything from wet sand over dry soil to
moist earth. There are over 100 different types of mines, even more
types are ``self'' made. We distinguish between Anti-Tank (AT) and
Anti-Personnel (AP) mines. In this stage of the project, we will focus
on Anti-Personnel mines. Their depth in the ground may range
between surface level down to 0.5 meter
.
Let's first look at the ideal case of a mine finding agent. The above requirements make it very difficult to define a universal robot that is able to move around autonomously in all possible terrains. The sensor returns a clear image of a ground cut, no matter what the soil conditions are. The optimal software would predict a mine with 100% accuracy and 0% failure rate.
The reality looks quite different. It is clear that a first design of a mine seeking robot can not handle all types of mines and all terrains. The sensor should also be optimized for a given soil type. Therefore, it makes sense to define an easier problem for the first version.
Another problem is the image acquisition and treatment. The ground radar images contain a lot of noise. The different soil layers, high moisture content, air bubbles and other soil features all impact the resulting ground image. A stone returns a clear signal, similar to the one returned by a mine [Fri95]. Considering that depth, dimension, weight and explosive content, all result in different responses of the radar signal, the complexity separating a mine from another object becomes obvious.