Features Selection for Sensor Fusion in a Demining Robot

Humanitarian demining is a dangerous and time consuming task which leads to more victims among deminers. Thus large efforts are being made worldwide in order to develop autonomous robots able to detect landmines without human participation. Any single sensor currently available for antipersonnel landmine detection can not provide the required detection/false alarm ratio, thus a combination of several sensors together with sensor fusion techniques is used. One of the main problems in the practical implementation of sensor fusion for a demining robot is the lack of experimental data for training the algorithms. Formally using a small training and evaluation set does not allow to take conclusions about the results of sensor fusion. This work presents a sensor fusion strategy which helps to overcome the problem of having few experimental data by creating an unified landmine signatures database. Using this database for sensor fusion in general and for features selection in particular is described. In addition to commonly used statistical features several new features are proposed which reflect the shape and the nature of the object. Moreover the designed ROIs extraction algorithm provides additional information allowing to improve features extraction. The results of features selection using a combination of Mutual information and Hausdorff distance are presented. Preliminary results of sensor fusion for pulsed and continuous metal detectors are also presented.