Dynamic Prediction of Parallel CA-CFAR Processing Performance on Large Datasets

Marcello G. Costa, Renato Machado, Fabio Bayer

DOI: 10.14209/sbrt.2023.1570923860

Evento: XLI Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2023)

Keywords: CA-CFAR runtime processing energy saving parallel process

Abstract

The constant false alarm rate (CFAR) detection theory maintains good decision performance over statistical interference. For imagery applications, this class of detectors increases the processing complexity due to the element-wise matrix multiplication performed for each pixel resolution. Hence, the largest collected dataset implies computational float-point operations with squared growth order, which end up compromising its performance under real-time requirements and power consumption constraints. To make the CFAR processing a workable solution for General Purpose devices, enabling it for several emergent embedded technologies, this article presents a dynamic prediction of parallel cell-average CFAR model with energy and runtime processing constraints. Distributing the processing through CPU and GPU cores, the convergence to higher energy save and lower runtime processing is achieved and measurable, with the discharge slopes as a function of element-wise matrix products dynamically predicted.

Download