Airborne learning algorithms: a case of a ground target multiple hypothesis tracker

In current practice there is a wide disconnect between increasing availability of sensor data and our ability to apply advanced learning algorithms to process that data in real time. While multitudes of very promising learning algorithms are developed in the research community, years and decades pass before these algorithms become implemented in embedded platforms. We describe a feasibility study of deploying a ground target multiple hypothesis tracker (MHT) in a system suitable for airborne deployment in platforms like JSTAR and the Global Hawk UAV. For a typical tracking scenario, an MHT algorithm requires more than 10 GFLOPs, necessitating implementation on a multicomputer. In addition, an MHT algorithm presents serious implementation challenges due to code complexity and global nature of the data. We discuss steps necessary to implement and deploy an MHT algorithm. We argue that modern high-performance computing systems that consist of general-purpose processors like Motorola PowerPCs and high-bandwidth interconnect like Mercury´s RACE++, coupled with advanced development tools and libraries, provide an environment that enables rapid deployment of advanced learning algorithms.