視覺監控資料錄影與重播之研究

Abstract

在本篇論文裡，我們提出了一個完整的即時監控系統的架構，除了包含了錄影、即時播放及重播等基本功能外，更針對壓縮儲存的方式提出了更進一步的探討。另一方面，由於攝影機的掃描方式為交錯式(Interlace)訊號，因此我們提出了一個去交錯(de-interlace)的方法，也針對了影像的縮放進行探討，希望能在重播時得到較佳的影像品質。 在壓縮儲存方面，我們採用軟體的壓縮方式，採取以MPEG4 simple profile為主體的壓縮方式。為了能使壓縮率更加提高，我們針對訊號與雜訊進行分析，在完全沒有訊號的情形下，採取Frame Skip的機制；而當有訊號發生的時候，我們設計出Block-Change的方式，根據訊號佔據畫面的大小來調整錄製的畫面；最後，當畫面進入長時間沒有訊號的情形下，我們會將調大MPEG4的QP值，讓壓縮率進一步的提高。在去交錯的部分，我們提出的方式主要的架構是以MA-IPC (Motion adaptive interlace-to-progress conversion)為主體，在空間域部分我們提出了一個新的方法，以分群為概念，綜合了LA(Line Average)以及Median的方式，進一步的提高去交錯後的影像品質。最後在影像縮放部分，我們根據影像內容，在平滑區與邊界區做不同的處理，綜合了NN (Nearest Neighborhood)、Bilinear以及Win-scale的方法，並且額外提出了一個model-based NN的概念，以達到影像縮放時較好的影像畫面品質。

In this thesis, a real-time surveillance system is proposed. Besides real-time monitoring, recording and replaying, we also focus on the enhancement of compression rate. Additionally, because the capture frames are in the interlaced form, we propose a new de-interlacing method. Moreover, we also propose a new image scaling method to improve the quality of scaled images. In compression and recording, we adopt a software solution that is based on MPEG4 simple profile. To further enhance the compression rate, three additional methods are adopted. If there is no change in the monitored scene, we perform frame-skip. If there are changes in the scene, we will detect the corresponding parts of the images and save only the information of those parts. Moreover, if the video has been under no change for a while, we’ll also change the value of QP (Quantization Parameter) to further enhance the compression rate. In replaying, a new de-interlacing method is proposed which is based on MA-IPC (Motion adaptive interlace-to-progress conversion). In the spatial domain, we propose a new method based on grouping. Based on grouping, the proposed method combines the advantages of LA (Line Average) method and the Median method. Moreover, a content-based image scaling method is also proposed which is a combination of NN (Nearest Neighborhood), Bilinear and Win-scale. We also propose a new idea named model-based NN to improve the quality of scaled images.