A Study on Unified Framework for Object-based Video Retrieval and Human Event Detection

瀏覽統計 Email 通知 RSS Feed

簡歷

基本資料

Project title

Code/計畫編號

NSC100-2221-E019-054-MY3

Translated Name/計畫中文名

無縫感知之視訊安全偵測、隱私保護與事件搜尋整合系統-以物件為基礎的視訊檢索、事件分析整合平台研究

Project Coordinator/計畫主持人

Shyi-Chyi Cheng

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=2854835

Year

2013

Start date/計畫起

01-08-2013

Expected Completion/計畫迄

01-07-2014

Bugetid/研究經費

597千元

ResearchField/研究領域

資訊科學--軟體

"大多數基於內容的視訊分析方法的失敗因素可歸咎於無法從雜亂背景檢測語義物件的準確結構或計算複雜度太大，為了解決這些問題，視訊導向應用常用的技術包括前景/背景分割、事先學習的活動、運動估計、或基於內容的視訊分析追蹤，然而，這些技術仍然是具挑戰性的研究議題。在本計畫中，我們提出一包括語意視訊片段切割、叢集處理、視訊物件檢索及特定事件偵測的整行性的方法，解決內容導向的視訊應用的相關問題，並建構基於視訊物件索引之特定事件辨識系統驗證本計畫所提出的整合性平台的正確性。本計畫分為三個部分，每部分都規劃為一年期的子計畫。首先，第一年的子計畫提出一種使用最大穩定影像區域(MSER)及區域霍福轉換投票機制切割資料庫視訊序列為若干具獨立語意的視訊片段，並據以擷取各視訊片段的重要視訊物件資料。本研究的目標是應用這些區域特徵探索視訊物件的結構，本方法克服視訊物件幾何變化及移動的問題。本計畫的第二部分提出一利用視訊物件特徵叢集方法索引視訊物件，並據以建立特定視訊物件檢索系統，當查詢視訊包含一視訊物件時，本系統可偵測出待比對視訊的所有匹配樣本。最後，本計畫的第三部分研究應用前兩年研究結果及視訊索引技術於特定事件偵測系統設計，本系統利用人類物件的運動軌跡偵測特定事件。總括一句，我們的系統中不需要執行一般視訊導向應用常用的技術包括前景/背景分割、事先學習的活動、運動估計、或基於內容的視訊分析追蹤，這些往往引發其他困難的問題降低系統效能。初步實驗結果驗證本計畫提出的方法的可行性。" "The advance of the content-based video analysis leads to mass video-based applications, such as video surveillances, video retrieval, video-based human activity detection and recognition, and sports analysis. However, most of the approaches for content-based video analysis either fail in detecting accurate structures of semantic objects from cluttered background or suffer from high computational complexity. In this project, we present a unified framework for specific event detection. The system combines shot detection, clustering, object detection and retrieval, and event detection using region-based Hough transform which groups local invariant regions into multiple objects. An application to human event detection is also constructed to verify the effectiveness of the system. The project is divided into three parts where each of them is scheduled to complete in a year. First of all, this project presents a shot segmentation based on a single set of maximally stable extreme regions (MSERs) using the voting method of the region-based Hough transform method. The important objects in every shot are then extracted. The objective of this research is to develop and apply computer vision methods that explore the structure of a video object. The proposed method is invariant to orientation, position, and motion. The second part of the project is to present an approach to object-based video retrieval using region tracks clustering based on object trajectories. The object indexing of database video sequences is established to construct an efficient system for video retrieval. Finally, the third part of the project is to present a human event detection system for video surveillance based on human object video indexing. The trajectories of video objects are also used to distinguish the difference among human objects. The generic techniques including foreground/background segmentation, prior leaning of activities, and motion estimation or tracking for content-based video analysis are not required in our system. Preliminary experimental results show that the proposed methods give good performance in terms of detection accuracy and robustness."

Keyword(s)

電腦視覺
事件偵測
區域霍福轉換
語意視訊片段切割
幾何轉換不變性
視訊物件檢索
Computer vision
event detection
region-based Hough transform
semantic-based shot segmentation
affine invariant
video object retrieval

DSpace CRIS

A Study on Unified Framework for Object-based Video Retrieval and Human Event Detection

基本資料

Description