Biologically motivated cross-modality sensory fusion system for automatic target recognition

Abstract

The need for robust target/background segmentation has led to the use of multiple band sensing systems. These sensors usually include some combination of visual, radar or laser range, and thermal infrared modalities. Despite over a decade of research, there are still a number of problem areas with existing automatic target recognition systems. Foremost among these are the high false-alarm rates frequently encountered due to nonrepeatability of the target signatures and possible obscuration of the targets from camouflage, environmental and sensor variations (Roth, 1989, IEEE Transactions on Systems, Man and Cybernetics, 19, 1210–1217; Roth, 1990, IEEE Transactions on Neural Networks, 1, 28–43). This paper presents a biologically motivated neural network system based on the rattlesnake that integrates multichannel sensory inputs for ATD/R. The system demonstrates a probability of detection greater than 90% with false-alarm rate less than l0⁻⁵ false-alarms/km² jor very small fixed targets using two-channel infrared input. In addition, temporal properties of the thermal neurons in the rattlesnake are demonstrated to be of possible use for segmentation of mobile targets from background clutter. Also presented are the results of some experimental studies on real-world multichannel infrared images sampled throughout a day.