Improving Mine Recognition through Processing and Dempster-Shafer Fusion of Multisensor Data

We propose in this chapter a fusion approach for improving anti-personnel mine recognition. Based on three promising and complementary mine detection sensors, we first address the issue of extraction of meaningful measures, which are then modeled as belief functions and combined within a Dempster-Shafer framework. A starting point in the analysis is a preprocessed multisensor data set containing some mines and false alarms. A method for detecting suspected areas is developed for each of the sensors. A detailed analysis of the chosen measures is performed for the selected regions. A way of including the influence of various factors on sensors in the model is presented, as well as the possibility that not all sensors refer to the same object. For every selected region, masses assigned by each of the measures and each of the sensors are combined, leading to a first guess on whether there is a mine or a non- dangerous object. An original decision rule adapted to this type of application is described too.