High Scalability Video ISR Exploitation.

摘要

The Intelligence Community uses computer vision (CV) algorithms on data from state-of-the-art sensors and platforms (e.g. Autonomous Real-Time Ground Ubiquitous Surveillance, ARGUS) on the National Image Interpretability Rating Scale (NIIRS) at level 6. Ultra-high quality cameras like the Digital Cinema 4K (DC-4K), which recognizes objects smaller than people, will be available soon for Wide-Area Motion Intelligence (WAMI). However, even if a platform was equipped with DC-4K, it would useless without CV algorithms to quickly process the vast quantities of data captured. (U) Today, several CV algorithms scale up by partitioning data across multiple nodes of computation. The standard approach is to increase the amount of processing power (e.g. GPUs) available. However, to achieve NIIRS level 7+ an emerging problem must be addressed: hard disk read latency. Reductions in disk read latency have not kept pace with increases in volume from 1990 through today. For example, reading a 3TB full disk from beginning to end takes hours. Considering that a NIIRS-8 DC-4K camera captures 32MB of data per frame, it would capture 3.2TB of data in a single mission hour. Industry has already solved this 'big data' problem in large-scale text processing through cloud computing architectures like Apache Hadoop. Hadoop applies a parallel batchprocessing paradigm that reads data from multiple hard disks simultaneously called Map/Reduce. In contrast to Hadoop, Modern CV algorithms assume a sequential data stream being read in order. While certain computations can be made in parallel already, integrating and correlating features computed simultaneously from randomly accessed portions of the data stream is an unmet challenge.