Multisensor fusion using neural networks

A multisensor system for robot navigation has been developed that uses artificial neural networks to perform sensor data fusion. Four neural networks were investigated regarding their potential to fuse data from an ultrasonic and an infrared range finder to yield more accurate estimates of depth. A radial basis predicter using localized receptive fields (LRF) was able to learn mappings quickly. However, it failed to locate receptive fields correctly within the input space thus providing a poorer mapping than a backpropagation network. A variation on LRF incorporating dynamic node creation was able to learn good mappings in about the same amount of time as a backprop network while exploring different network sizes. An output encoding scheme produced the best performance by exhibiting less error at places where the depth functions that varied rapidly. The resulting network provides a cost-effective solution to range estimation for autonomous navigation using on-board hardware