https://scholars.lib.ntu.edu.tw/handle/123456789/174881
標題: | 利用視覺導向影像穩定技術之演算法與硬體架構設計 Visual-Centric Algorithm and Architecture Design of Video Stabilization |
作者: | 黃耿彥 Huang, Keng-Yen |
關鍵字: | 影像穩定;硬體架構;video stabilization;hardware architecture | 公開日期: | 2011 | 摘要: | 本篇論文主要在探討如何解決影像晃動的問題。如何解決影像晃動的方式便是所謂影像穩定演算法所希望達到的目標。影像穩定系統最主要的目標是希望能將晃動的影像分析出非預期震盪和有意移動,而透過補償非預期震盪,我們可以得到一個只存在有意移動的穩定影像。傳統上,為了減少因為攝影機的晃動造成影像不穩而產生的副作用,我們會將影像擷取裝置放置在一些固定的位置上,比方三腳架或者攝影機的滑軌。然而,隨著科技的發展,影像擷取裝置愈做愈小,這些攝影機變成容易攜帶在各種不同的活動載具上,影像不穩定的現象就變得愈來愈明顯。而未來的趨勢是每個移動攝影機上的功能也會愈來愈多元,而不穩定的影像會造成往後在處理上的困難,這時對於影像穩定的需求變愈來愈迫切。為了解決這個惱人的問題,電腦視覺的各項演算法可以幫助我們直接分析得到的影像,我們便可計算出攝影機的晃動狀況以及該如何補償這些不穩定的畫面。不過還有很多問題有待解決,這些電腦視覺的演算法往往需要相當長的計算時間才能提供一個穩定的答案,這樣便大幅限制影像穩定系統只能當成是後處理的系統,也就是從已經錄製好的影像裡去分析並重新建出一個穩定的影像。然而在高科技的未來,及時運算的重要性將會愈來愈大。 我們在這篇論文中提出一套高效率的影像穩定系統,目的是希望能夠提供一個即時的影像穩定系統當作是一個更大更複雜系統的前處理工具,幫助不同的影像處理演算法都可以因為有高品質穩定影像的輸入,整體的系統效能可以大幅的提升。為了達成這樣的目標,我們提出了以視覺導向為基礎的影像穩定演算法並且符合及時運算的條件,接下來將介紹一些此演算法的特色。核心的概念是不同的東西應該有不同的價值,而我們應該要將重心放在我們真正關注的東西上。本篇論文採用視覺鎖定(Visual Fixation)的概念,不同的特徵點(Feature Point)會根據個別的重要性和信賴程度給予不同的比重。為了更快速的處理影像,特徵點的比對非常重要,因此我們提出使用高效率的空間敏感散列(Locality Sensitive Hashing),並且再透過硬體加速,達到快速的比對。而硬體加速的方式主要是透過記憶體的減少以及利用暫存記憶體增加特徵點的比對速度也減少不必要的特徵點傳送。 藉由我們提出的影像穩定系統方式,在不同的測試環境下依然可以達到平均90%以上的準確度,而晶片的記憶體大幅減少了90%。除此之外,重複的特徵點搜尋降低了50%以上。透過以上的各種演算法改進以及硬體加速,本篇論文最後實作出一個完整的影像穩定晶片,在UMC 90奈米的製程下,晶片面積為3mm x 3mm,記憶體使用了41.54KB。支援影像解析度達1280x960,並且畫面的頻率平均高達202fps。 This thesis is to investigate the need for solving the problem of shaky video stream. The way we try to stabilize the video is also known as the video sta- bilization system. This stabilization system is for removing the unwanted oscillation while conserving the intentional camera motion. In the past, people tend to resort to he help of some devices like tripod to reduce the negative eRects induced by unstable camera platforms. However, with the development of camera image retrieving technology, camera sensors are be- coming more and more small that we have to think other ways to stabilize the annoying video stream. By the assistance of computer vision technique, there is information that we can learn directly from the video itself. Thus, various computer vision based approaches have been brought up to com- pensate the uncomfortable video sequence. Unfortunately, these computer vision based approaches often require longer processing time to guarantee a certain quality. As a result, many algorithms are utilized in post processing stage to reexamine the shaky video and reconstruct a new stabilized result. But there is more need for a real-time system that can provide a smooth video result. An e±cient video stabilization system which is targeted at real-time applications is revealed in this thesis. For various applications or further image processing techniques, performance can be largely enhanced if pro- vided with a video in high quality. In order to achieve this goal, we develop a visual-centric video stabilization algorithm. The core idea is to provide a real-time stabilization technique for applications, thus, we propose a high e±ciency video stabilization system with the following special characteris- tics. The  ̄rst and the main idea is that we should put more emphasis on the information that we really care about. Visual  ̄xation is thus adopted in our system. Each extracted feature is weighted diRerently based on its visual importance and reliability. In order to achieve real-time processing target, high e±ciency feature matching engine is constructed in our system. Locality sensitive hashing helps us to realize the feature matching procedure in hardware implementation. By applying the proposed bucket memory re- ducing scheme, the memory consumption of locality sensitive hashing is reduced by more than 90% compared with allocating full size of hash ta- bles. For insu±cient throughput, we adopt a 2 way set-associative cache to reduce the frequency of accessing outer feature memory. More than 50% of feature bus burden has been released. Through our proposed locality sen- sitive hashing scheme, we make it possible for the digital signal processing based real-time video stabilization system. The experimental results show our improvement and robust performance under diRerent conditions. The proposed video stabilization system is realized on a chip with 3£ 3 in area using UMC 90nm technology. Average 202 frames per second of processing time is achieved |
URI: | http://ntur.lib.ntu.edu.tw//handle/246246/256771 |
顯示於: | 電子工程學研究所 |
檔案 | 描述 | 大小 | 格式 | |
---|---|---|---|---|
ntu-100-R98943037-1.pdf | 23.32 kB | Adobe PDF | 檢視/開啟 |
在 IR 系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。