In the mid and late 1990s Data Fusion was regarded as potentially a key method for both improving the detection rate and especially reducing the false alarm rate of mine detectors. However, it has proved substantially more difficult to implement than was originally envisioned, partly due to the heterogeneous nature of the sensors. Currently (2004/5) the few hand-held dual sensor mine detector systems under test or deplyment make little use of data fusion and instead prefer to use metal detection as a primary sensor and GPR as a secondary method for confirmation. Little information is available about state-of-the art vehicle mounted multisensor systems for military demining.

The principle of data fusion is that, by combining either data or information from complimentary sensors (i.e. sensors using different physical principles such as metal detection, GPR, radiometers, vapour detection, etc), detection can be improved and false alarms reduced. There are a number of different approaches which can be employed, and selection of the correct approach and algorithms is not necessarily straightforward.

The choice of a method for merging data in a multi-sensor system depends on the data format of the signals, which can be a 2D image (e.g. IR), a 1D time series (e.g. GPR), or a scalar value which expresses the detection of the presence of explosive or metal for example. Three different types of data fusion architectures can, in principle, be used:

• Pixel level fusion: multiple images are combined to a single image, and each location in the combined image has an associated vector of measurements from each of the sensors. The new image is then processed by an algorithm such as target detection/recognition that simultaneously operates on the vector values. The problems of putting pixel level fusion into practice are due to the differences in field of view, in sensor orientation (e.g. forward looking, downward looking), in resolution, and in data format.

• Feature level fusion: features are extracted from each of the sensor data, followed by a registration step, usually carried out at the level of regions of interest or image segments containing more than one pixel. Such a co-registration of features from individual sensors is often easier to achieve than pixel level fusion. A detection/classification algorithm can then be applied on the combined feature vector characterising a region of a certain spatial extent.

• High-level data fusion: each sensor makes an independent decision based on its own observations and passes these decisions to a central fusion module where a global decision is made. Because the sensors have very different data characteristics, this kind of data fusion is probably the most accessible for a mine detection system.

A number of projects have developed and implemented data fusion techniques but as yet there is little or no field experience of DF for mine detection. However, the EU LOTUS project successfully demonstrated a proof-of-concept vehicle-mounted multisensor system (MD, GPR and IR camera) using real time data fusion based on modified Bayesian algorithms. This is believed to be the first such demonstration, for Humanitarian Demining, in Europe. The HOPE and DEMAND projects also developed data fusion techniques. The HOM-2000 project in the Netherlands has scheduled data fusion for handheld and vehicle based systems in its starting second phase.