請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/44385
標題: | 適用於智慧型攝影機系統之行為分析演算法與架構設計 Algorithm and Architecture of Behavior Analysis for Intelligent Camera Systems |
作者: | Tzu-Heng Wang 王子恆 |
指導教授: | 陳良基 |
關鍵字: | 行為分析,硬體架構,粒子濾波器, Behavior Analysis,VLSI,Particle Filter, |
出版年 : | 2009 |
學位: | 碩士 |
摘要: | 本篇論文中提出了一個用於智慧型攝影機系統的行為分析平台。從最近的研究走向來看,影像分析在智慧型攝影機系統中扮演重要的角色,而且可以用於像是智慧型監控系統、醫療系統、或是人機介面等等,而從影像中的行為分析將為未來攝影機系統一個重要的功能。因此,我們的設計以行為分析演算法和架構為出發點,並使用無手把的遊戲動作為我們的假想驗證環境。
我們所提出的行為分析平台為高度整合化的平台,能達到「影像進、動作出」的整合化功能。演算法方面,我們支持並實驗三種不同的行為分析演算法,這三種演算法隸屬在兩大類的行為分析演算法底下,分別為動態分析和全面式分析。動態分析結合了追蹤模組和軌跡分析模組,追蹤模組負責產生身體關節點的位置和移動軌跡,而軌跡分析模組則將這些產生的軌跡轉為行為的表示方法。全面式分析則有時序模板演算法和運動向量模型兩種。我們的實驗結果為動態分析能得到92.57%的辨認結果,為三者中最好的,而另兩個演算法的辨認率分別為運動向量模型的92.38%和時序模板的81.37%。 即時的效能表現對於行為分析演算法亦是一項重要的要素。因此我們將所支持的演算法作了效能和執行時間分析,發現追蹤模組占了系統中超過90%的運算量,而且在軟體上只能達到每秒處理5~6張影像的速度。為了滿足即時運算的需求,我們提出了追蹤演算法,也就是粒子濾波器的硬體加速架構。設計的同時,不同的最佳化技巧亦被應用在硬體的架構上。在使用硬體加速之後的系統可達每秒34.687張影像的運算速度,因此能達到即時運算的需求。我們最後的硬體實作於TSMC 90nm 1P9M Low-k Logic Process製程,面積為2.221 x 2.101 mm2,最高運作頻率為125MHz,所需的功耗為185.2mW。 Video-based analytics is a key component of intelligent camera systems to fulfill applications such as intelligent surveillance, healthcare and human computer interface in the future. In this thesis, a behavior analysis platform for intelligent camera systems is proposed. We motivate our design by the needs of human behavior analysis among different scenarios and use controller-free gaming system as the demonstration. The proposed human behavior analysis system is fully integrated, in which we integrate the tracking module and trajectory analysis module, supporting the functionality of “video in, action out”. Three behavior analysis algorithms are supported in two different major approaches of behavior analysis, that is, the dynamics modeling approach and the holistic approach. The tracking algorithm based on particle filters can estimate objects’ positions, sizes and angles, and the body joint trajectories generated by tracking are modeled and analyzed to action descriptors. The recognition rate of the dynamics modeling approach can achieve 92.57% in average. The proposed tile-based motion vector pattern algorithm and temporal templates algorithm are supported as holistic approaches. The tile-based motion vector pattern algorithm can achieve 92.38% recognition rate by averaging all cases, and the temporal templates algorithm can achieve 81.37% in average. Real-time performance is needed for the behavior analysis system to be integrated into cameras. The run-time analysis and profiling results shows that the tracking module requires most of the computation, thus the hardware architecture of color-based particle filter is proposed. Optimization techniques are utilized when designing the hardware. Tracking can operate at the normalized frame rate of 31.35 F.P.S. by the hardware, and the run-time of the behavior analysis system is at least 34.687 F.P.S., achieving realtime performance. The final chip is implemented in Verilog HDL and TSMC 90nm logic process is the fabrication technology. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/44385 |
全文授權: | 有償授權 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-98-1.pdf 目前未授權公開取用 | 18.19 MB | Adobe PDF |
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。