Usage of polarisation features of landmines for improved automatic detection

In this paper the landmine detection performance of an infrared and a visual light camera both equiped with a polarisation filter are compared with the detection performance of these cameras without polarisation filters. Sequences of images have been recorded with in front of these cameras a rotating polarisation filter. Due to optical distortion of the rotating filter the sequence contains apparent motion. This motion is estimated from the images and corrected using bilinear interpolation. From the motion corrected sequences both the unpolarised intensity and the polarisation parameters are estimated. The detection performance is evaluated with Receiver Operator Characteristics curves. It is shown that, for higher detection rates, classifiers based on polarisation features perform better than classifiers based on the intensity cue only. The linear combination of both has the best overall performance for the scenes considered. The infrared camera outperforms the visual light camera in the difficult scene.