Hardware Enabled Acceleration of Near-Field Coded Aperture Radar Physical Model for Millimetre-Wave Computational Imaging

Abstract

There is an increasing demand in real-time imagery applications such as rapid response to disaster rescue and security screening to name a few. The throughput of a radar imaging system is mainly controlled by two parameters; data acquisition time and signal processing time. To minimize the data acquisition time, various methods are being tried and tested by researchers worldwide. Among them is the computational imaging (CI) technique, which relies on using coded apertures to encode the radar back-scattered measurements onto a set of spatio-temporally incoherent radiation patterns. Such a CI-based imaging approach eliminates the requirement for a raster scan and can substantially simplify the physical hardware architecture. Equally important is the processing time needed to retrieve the scene information from the coded back-scattered measurements. In CI, the simplification in the hardware layer comes at the cost of increased complexity in the signal processing layer due to the indirect mapping and compression of the scene information through the spatio-temporally incoherent transfer function of the coded apertures. To address this particular challenge, this paper presents a hardware-based solution for CI signal processing using a Field Programmable Gate Array (an Xilinx Virtex-7 (XC7VX485T) FPGA chip) architecture. In particular, the proposed method consists of calculating the CI sensing matrix using the FPGA chip and storing it on the FPGA platform for image reconstruction. For the adjoint operation, the calculated sensing matrix is applied on the measured back-scattered waves from the target object. We demonstrate that the FPGA based calculation can reach 21.9 times faster speed than conventional brute-force solutions.

Fourier Compatible Near-Field Multiple-Input Multiple-Output Terahertz Imaging With Sparse Non-Uniform Apertures

“…All results are provided for signal-tonoise ratio (SNR) of 20dB to ensure a realistic channel response in the presence of additional loss factors. Our previous works in NF imaging [34][35][36] justify this SNR selection. According to the parameters of Configs.…”

Coded-Aperture Computational Millimeter-Wave Image Classifier Using Convolutional Neural Network

“…(1), f signifies the reflectivity distribution of the discretized pixels in the imaged scene, H is the sensing matrix and n is the measurement noise, modelled as a Gaussian distribution with zero mean [33]. In this work, to ensure a realistic model, the back-scattered measurements, g, exhibit a finite SNR of 20 dB [33], [34]. Subscripts M and N denote the number of measurement modes and the number of pixels used to define the scene, respectively.…”

“…In a typical CI setup, which uses multiple apertures and usually involves imaging a large scene, it has been shown in the past that processing of sensing matrices as large as 90 GB can be required [35]. Such a calculation tend to be intensive computationally, hence an FPGA architecture is employed in the reconstruction process to share the computational load with the CPU, as presented in [34]. This hardware architecture makes use of the FPGA logic blocks to carry out the calculation of the sensing matrix in quasi realtime.…”

Fourier-based Radar Processing for Multistatic Millimetre-wave Imaging with Sparse Apertures

“…To alleviate the above constraints, computational imaging approaches have been used in the literature [4][5][6][7]. However, such configurations demand the inversion of an ill-posed matrix, which results in a bottleneck in achieving fast image reconstruction time.…”