Sensor Fusion Implementation with Neural Networks and/or Fuzzy Logic

Today's sensor data fusion approaches typically use some form of an extended Kalman filter to fuse the data. A basic assumption is that the sensor data is contaminated by additive white noise disturbances. This is far from being true with some of today's sensors - considering their operation nonlinear regimes: saturation, deadzone, distortion, etc. In this proposal, we plan to account for these sensor nonlinearities and uncertainties using a fuzzy logic/neural network approach. Fuzzy logic is used to account for the nonlinear sensor uncertainty via multiple membership functions for a candidate set of sensor systems. The neural networks are then used to operate on the sensor data to exploit relationships between the sensors - this significantly enhances the data fusion task. We propose to use ORINCON's experience in multiple-hypothesis tracking to conceive of virtually every permutation of how sensor data can be fused from within a multitarget scenario. The neural-fuzzy networks ar inserted to account for the sensor nonlinearities, to intelligently select the proper sensor combinations, and to fuse the data to arrive at the most accurate target definitions (localization and classification). The blending of the multiple-hypothesis tracking architecture with the neural-fuzzy networks is a computationally efficient and cost-effective approach for sensor data fusion that could lead to more accurate target scene descriptions.