Research And Implementation Of Key Technologies In Single-pixel Video Sampling Based On Compressive Sensing

Based on Nyquist-Shannon sampling theorem, classical sampling methods convert analogsignals to digital signals. However, these methods cause deluge of data at the same time.Compressive sensing(CS) is a novel sampling theory different from Nyquist-Shannonsampling theorem. It can accurately reconstruct original signals after sampling signals usinglow sampling rate that is below Nyquist rate. Therefore, it receives broad attention fromresearchers in different fields. Using CS theory, researchers in Rice University successfullyimplement single pixel camera which linearly measures the light of scene with only onephotodiode. On the basis of single pixel camera, researchers begin to study the schemes of CSvideo sampling. But many current video sampling methods acquire low-quality video andsuffer from high time cost. This thesis deeply studies the CS theory and the framework ofsingle pixel camera. Then we analyze the CS video sampling methods and propose animproved algorithm in order to address the problems mentioned above and increase the videosampling efficiency. The main contributions of this thesis are as follows:1. To better understand the principle of single pixel camera, we set up the single pixelcamera experiment platform and study the relevant reconstruction algorithms such asTVAVL3 and OMP. This platform verifies the feasibility of single pixel camera and providespreparation for following sections.2. To enhance the video quality, a multi-frame motion estimation algorithm is proposedand used in the CS video sampling method. Instead of using two frames motion estimation,the proposed algorithm uses multiple frames to implement motion estimation. Experimentsshow that this algorithm can improve the quality of recovered videos.3. To reduce the motion estimation time, an alternative algorithm is proposed to replaceoptical flow computation. In the proposed method, block match algorithm is used to processmotion estimation. Experiments demonstrate that even though it slightly influences the qualityof recovered videos, it can vastly reduce motion estimation time.