Design of a Power Efficient Accelerator for Reconstructing Videos from Gaussian Mixture Model Data

The Gaussian Mixture Model (GMM) is a compressive sensing (CS) technique which can effectively reduce the sizes of video files. However, reconstructing a GMM-based CS video is very time-consuming because it usually involves a process with high computational complexity. In this paper, we re-design the reconstruction flow to make it hardware friendly and implement a power-efficient accelerator based on this modified flow. The proposed accelerator, with the power consumption of only 1.5 Watts, can complete reconstruction tasks 25.07 times faster than a server-class CPU for a CS stream with a resolution of 300×400 pixels.