SENSOR-FUSION IN NEURAL NETWORKS

摘要

The problem of sensor-fusion arises in many applications. We have studied the problem primarily in the context of cognitive robotics. An autonomous robot has to create a meaningful representation of its position in the world and, more generally, the situation in which it has to perform an appropriate action. Such robots usually have several different sensors (sonars, laser range finders, cameras, microphones, infrared sensors, bumpers) and they have to combine the different kinds of information provided by these sensors. Neural networks seem to be particularly well suited for the combination of inputs from completely different sources. So we will be using them and the corresponding learning and training strategies for these problems. In the context of pattern recognition, a community has emerged that is analyzing these problems and has accumulated a considerable amount of practical and theoretical knowledge in the theoretical framework of multiple classifier systems. There is a strong overlap between this community and our neural networks community, and we have also contributed to the application of multiple classifier systems built from neural networks for sensor-fusion, for example, for the classification of multivariate biological time series. One question to be addressed here is the use of uncertainty measures for the combination of classifiers. Here we use the framework of multiple classifier systems to compare several methods of sensor- or information-fusion.