Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia. E-mail: davor.antonic@fsb.hr.
Language: Croatian.
Supported by: EPFL, Croatian Ministry of the Interior and CROMAC.

ABSTRACT

This thesis describes development and implementation of algorithms for signal processing, feature extraction and classification, as well as their integration into the system for buried landmine detection. Procedure for automatic generation and reduction of the initial feature set directly from signal spectrums has been developed. Novel feature extraction algorithms based on heuristic tree search and genetic algorithms, which are applicable to the wide range of problems have been proposed.

All the algorithms were tested on real mines and other objects. Application to the set of sonar samples resulted in significant reduction of classification system complexity. Algorithms were also tested on mines and other objects samples. Samples were acquired by developed detector, based on standard deminer`s prodder which records vibrations produced by touching the buried object. Testing shows that proposed detection system and dedicated set of procedures make mines distinguishable from other buried objects.

Appropriate information support has been defined, which makes possible managing information created at the level of individual mines detection. It allows incorporation of the information into the humanitarian demining information system and establishment of feedback which will make possible learning and continuous improvement of detection system features. Methods for position tracking of the detector attached to the remotely controlled or autonomous robotic platform are also described.

Keywords: Digital signal processing, feature extraction, heuristic tree search, genetic algorithms, pattern recognition, neural networks, information systems, mobile robots, demining.