Fusion of polarimetric infrared features and GPR features for landmine detection

Currently no single sensor reaches the performance requirements for humanitarian landmine detection. Using sensor-fusion methods multiple sensors can be combined for improved detection performance. This paper focuses on the feature-level; fusion procedure for a sensor combination consisting of a polarimetric infrared imaging sensor developed by TNO and a video impulse GPR developed by Delft University of Technology. Feature-level sensor-fusion is the process where specific information (i.e. features) from objects detected by different sensors are combined and classified. The single sensor detection methods and the feature-level sensor-fusion methods are evaluated using a leave-one-out evaluation method. This evaluation method provides an independent evaluation set while retaining the largest possible training set. The detection results of both single sensors and the sensor-fusion methods are both presented in receiver operator characteristics (ROC) curves. They show that on the training set feature-level sensor-fusion always outperforms the best single sensor. Furthermore, on the independent evaluation set, there are ROC points of the feature-level sensor-fusion methods that are better than the best sensor.