Sensor Fusion for Hand-held Multi-Sensor Landmine Detection

Sensor fusion issues in a streamlined assimilation of multi-sensor information for landmine detection are discussed. In particular multi-sensor fusion in hand-held landmine detection system with ground penetrating radar (GPR) and metal detector sensors is investigated. The fusion architecture consists of feature extraction for individual sensors followed by a feed-forward neural network training to learn the feature space representation of the mine/no-mine classification. A correlation features from GPR, and slope and energy features from metal detector are used for discrimination. Various fusion strategies are discussed and results compared against each other and against individual. Both feature level and decision level fusion have been investigated. Simple decision level fusion schemes based on Dempster-Shafer evidence accumulation, soft AND, MIN and MAX are compared. Feature level fusion using neural network training is shown to provide best results. However comparable performance is often achieved using decision fusion based on Dempster-Shafer theory. It is noted that the simple fusion scheme lacks a means to verify detections after a decision has been made. New detection algorithms that are more than anomaly detectors are needed. Preliminary results with features based on independent component analysis (ICA) show promising results towards this end. An improved fusion scheme using ICA features is proposed