Combining rule-based and neural-network approaches in the fusion of range and intensity signals

This paper exposes our approach to Sensor/Data Fusion through a prototype system developed for feasibility study and experimentation. System specifications include the fusion of data from multiple sensory modalities, processing in layers of abstraction, integration of qualitative and quantitative information processing, combination of alternative processing paths, and capabilities for learning. The domain selected is that of fusing a light intensity signal with a registered range signal. The system goal is to determine the existence and location of edge patterns; these correspond to occluding edges, surface orientation change edges, and edges denoting reflectance differences. The system combines two alternative processing paths, a feature-layer fusion process which uses classification rules, with a signal-layer fusion process which utilizes a Radial Basis Functions neural network. While the former process exploits predefined rules, the latter is able to learn from examples. Experiments show that, while any of the two alternative approaches can be invoked, their combined utilization is able to improve system performance