Land Mine Detection using Dual-Band Electro-Optical Sensing
This thesis addresses the detection of buried and partially occluded land mines using electro-optical sensing. According to the Red Cross approximately 10000 people are killed and another 30000 are injured each year in mine related accidents. In addition, the mine situation often prohibits economic growth in countries affected, as agricultural areas cannot be used and infrastructure cannot be developed.
Detection of buried land mines using electro-optical sensing is possible mainly by exploiting the infrared wavelengths. The buried mine disturbs the natural heat and mass transfer in the surrounding soil and leads to a thermal signature at the surface that might be possible to detect with an infrared camera. However, it is problematic to design a detector as the signature is weak and distorted by noise and clutter stemming from irregularities in the soil and at the surface. It is identified that the surface clutter, which is the prime concern, is highly correlated to the reflected surface illumination. By using an additional sensor operating in the visual wavelengths, the dual-band data can be exploited to enhance the detection performance.
Deriving detectors for such a scenario implies two main problems. Neither the mine signature nor the statistics of the background noise/clutter are known a-priori. The main theme of the thesis is to consider both parametric and non-parametric solutions to these problems. Specifically, a Bayesian perspective is promoted. Considering the problem of an unknown signature, the Bayesian approach can be used to incorporate prior knowledge regarding physical properties, such as recent weather and time of burial, into the detector. The Bayesian perspective is also studied in order to handle the problem of an unknown noise covariance matrix. By properly take into account the uncertainties regarding the noise color, an enhanced ability to describe the desired signal is obtained.
Finally the problem of detecting partially occluded surface-laid mines is considered. In such a scenario, the traditional additive clutter model is no longer applicable. Instead a more realistic model is presented and detectors are derived. The methods proposed in this thesis are evaluated both on synthetic data and real measurements and show promising results.