Edge-Based Machine Intelligence Architecture for In-Situ Video Processing using Binarized Neural Networks
Abstract
As new unmanned devices with enhanced features for reconnaissance and support are developed, the massive quantities of data that they transmit have necessitated new machine learning methods and computing platforms to perform higher-level reasoning computations at the edge of the computing network. This proposal responds to such challenges by advancing edge-based machine intelligence. Edge-based machine intelligence offers transformative and impactful alternatives to collect and transmit vast quantities of uninteresting raw data collected by unmanned devices. Thus, a focus and urgent need of current machine intelligence techniques have been toward more capable and efficient deep neural network algorithms to transform vast amounts of real-time data from the field into usable, decipherable, and actionable information in-situ. A proven and increasingly-capable class of niachine intelligence techniques for such applications are based on various Deep Neural Network (DNN) approaches. However, DNN algorithms are computationally-demanding and thus only feasible to operate primarily on powerful multi-core processors, GPUs/TPUs, and centralized accelerators, especially for real-time or near real-time operation. Simultaneously, while being powerful, advancements for in-situ deep neural network methods must also be extremely efficient and leverage new optimizations for restricted and finite energy budgets available on remotely-deployed robotic devices. Ideally, new edge-capable DNN algorithms would increase autonomy significantly, while also operating on miniaturized, adaptable, and tamper-resistant computing platforms, such as those realized by Field Programmable Gate Arrays (FPGAs). To address challenges ofDNN processing and analysis ofremotely-gathered real-time imagery and video, the Edge-Based Machine Intelligence Architecture for In-Situ Video Processing using Binarized Neural Networks project proposes to design, prototype, and refine novel neural network frameworks which minimize hardware resources, energy consumption, and execution time for edge-of-network processing. Innovative algorithm refinements, hardware implementations, and cross-layer optimizations will result in FPGA-prototyped frameworks demonstrating a reduction in throughput degradation, runtime latency, area cost, and energy consumption using currently-available FPGA synthesis toolchains. Difference detection and cost-based optimization algorithms, prototype frameworks, and test benches will be developed and used to conduct a systematic analysis of selected real-time video sequences. The project s deliverables will be evaluated using a validation case study with available benchmarks. The methods and platform prototype will advance the Army s mobile and Unmanned Aerial Vehicles (UAVs) capabilities for extracting useful information in real-time from remote locations and complex urban environments, thus leading to the more capable battlefield and training support.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 09, 2020
- Source ID
- W911NF2010174
Entities
People
- Yu Bai
Organizations
- Army Contracting Command
- United States Army