| Compressed sensing is a novel technology to acquire and reconstruct sparse signals below the Nyquist rate. This thesis explores the temporal redundancy in videos, and proposes a block-based adaptive framework for compressed video sampling. The proposed framework classifies blocks into different types depending on temporal correlation, and adjusts the sampling and reconstruction strategy accordingly. This framework also considers the texture complexity of regions, and adaptively adjusts the number of measurements collected. A frame rate selection module is included to select the maximum achievable frame rate under the hardware sampling rate and the perceptual quality constraints. Simulation results show that compared to the raster scan, the proposed framework can increase the frame rate by up to six times depending on the scene and the video quality constraints. A 1.5 ∼7.8dB gain in the average PSNR of the reconstructed frames is observed when compared with prior works on compressed video sensing. |
