Adaptive sensor data fusion architecture for landmine detection and discrimination

The aim of this paper is to develop a framework for multi-sensor data fusion for the detection and identification of anti-personnel mines as a part of humanitarian demining project. A two-stage hybrid architecture is proposed to integrate non-homogeneous and dissimilar sensor data from various sensors being developed as a part of the project. The first stage is used to extract significant information from individual sensor data. Self-organizing neural networks are used to define natural and significant clusters embedded in the sensor data. In this regard two popular self-organizing NN architectures of ART2 and DigNet are studied. The second fusion stage is used to integrate this local sensor information into a global decision. The global decision could be binary as in mine/no-mine decision set, or it could be more complex where identification of the underground mine may be involved. For the present paper, reliable data from different sensor was not available. Extracting different shape features like moment invariants and Fourier descriptors simulates dis-similar sensor data for simulated shapes. Some results for the performance of the clustering algorithms and the fusion architecture are presented.