SAR/IR Sensor Fusion Algorithms for Automatic Target Recognition

Abstract

Automatic Target Recognition (ATR) is an algorithm that is capable of detecting and recognizing the potential targets of interest using acquired sensor data. In both military and commercial applications, ATR is the key technology of the defense systems and is widely used in various autonomous missions. With the advance of computer vision, ATR technologies have made great progress in the last couple of decades. However, most ATR methods still utilize only single-sensor data and suffer performance degradation due to the uncertainty of data. Recently, a number of sensor fusion algorithms for ATR have been developed and achieved great improvement in the reliability of recognition results, but there is still a lack of research about the fusion of Synthetic Aperture Radar (SAR) and Infrared (IR) data. In this dissertation, SAR/IR fusion algorithms for ATR are presented. First, the decision-level SAR/IR fusion algorithm for target detection is introduced. The developed fusion method exploits Dempster-Shafer evidence theory to combine the detection results of SAR and IR. In addition, the detection results of each single-sensor data are integrated to remove the false alarms that are not filtered by Dempster-Shafer evidence theory. In target detection experiments, target detection rate and false alarm rate of the developed fusion algorithm are evaluated and compared with those of the conventional algorithms. Next, the sensor fusion algorithm using deep learning for SAR/IR target recognition is presented. Conventional deep learning algorithms for sensor fusion rely on simple fusion architecture and utilize pre-trained single-sensor streams that are trained independently without considering the fusion. The developed deep fusion architecture reduces the information loss caused by the defect of a conventional fusion network, and each single-sensor stream is trained to learn the complementary features using dissimilarity regularization. To prove the superiority of the developed fusion algorithm, the target recognition rate of the developed fusion scheme is compared with that of conventional fusion approaches. In addition, the effects of advanced fusion architecture and regularization are fully analyzed.