Infrared-based land mine detection on a vehicle

Abstract

The paper discusses a new method of land-mine detection with a multi-stage algorithm for an infrared (IR) sensor mounted on a vehicle. In recent years, IR-based detection systems have gained increasing interest for the land-mine detection problem. We propose a novel processing technique for the IR sensor. The method consists of five algorithmic stages: (1) pre-processing, (2) preliminary detection, (3) feature extraction, (4) classification, and (5) combination of multiple classifiers. The pre-processing step uses a novel adaptive filtering technique that enhances the mines with respect to the background by an online algorithm. The pre- processed IR image is analyzed by parametric and non- parametric methods of testing differences of means for populations from two distributions. We identify candidate regions of interest at the preliminary detection stage, from which we extract features for classification. We use three different classifiers which are based upon a (1) probabilistic neural network, (2) decision tree, and (3) multi-layer feed- forward neural network. Results from multiple classifiers are combined using a new technique of dynamic classifier selection. We apply the complete algorithm on images acquired by IR cameras mounted on a vehicle, and the preliminary test results are very encouraging.