Connectionist approach to the fusion of three-dimensional, sparse, unordered sensor data

A key component in the development of flexible automation systems is the sensor capacity of the system. When information is gathered from multiple sources, both quantitative and qualitative improvements in the data may be realized. In this paper we examine the sensor fusion problem which accompanies the use of multiple information sources. A modification of Kohonen's neural network algorithm is proposed. This modified approach, called the Doubly Constrained Neural Network (DCNN) learns the topological relationships present in the data while overcoming some of the problems found when using Kohonen's approach (i.e. the poor quality of fit in the presence of discontinuities). Results for several sets consisting of scattered, unordered, three-dimensional surface data, demonstrate the excellent topological fit achieved by the DCNN in cases where the Kohonen method fails. The method presented here shows great potential for enhancing the decision making capability of humans and machines which depend on multiple sensory sources or moving sensors for their perception of the environment around them. Potential applications include: part inspection for manufacturing, learning and identification of landmarks for mapping and navigation, and collision avoidance for maneuvering vehicles.