Classification of Radar Targets Using Neural Networks on Systems-on-Chip
Download
Date
Authors
Type
Examensarbete för masterexamen
Model builders
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The fast paced nature of modern combat situations has increased the need for quick
and adaptable radar classifications to identify potential threats. One potential option
to increase accuracy and performance is to introduce machine learning in radar
classification tasks. Utilising convolutional neural networks to identify patterns
within radar data provides an additional stream of information that could be used
to classify targets. These neural nets require high-performance processing whilst
still conforming to the low power and mobility requirements inherent in defence
applications. These requirements make FPGAs a natural choice to be used as a
hardware platform in radar classification tasks.
This project explores the capabilities of the new Xilinx Versal VCK190 ACAP which
combines regular FPGA architecture with AI Cores, which can be used to accelerate
neural network tasks. Our findings show that the amount of radar classifications
per second can be increased by at least 20x compared to a neural net running on
a consumer grade CPU. This increase was achieved by utilising the low latency
interfaces and high performance acceleration of the AI cores which are unique to the
new Versal platform. These aspects make the VCK190 an interesting platform to
further develop upon but more research needs to be made to improve the accuracy
of the model.
Description
Keywords
Xilinx, Versal, VCK190, Convolutional neural networks, Radar classification, FPGA, Thesis