以物件為基礎監控視訊內容之追蹤與摘要

Abstract

影像內容摘要是藉由擷取影像資料之中的重要涵義來精簡地表示影像資料之重要的內容。而在監控視訊，我們可以用以物件為基礎之視訊摘要來表達物件所發生的事件與其代表重要的意義。因此，在視訊監控系統如智慧型運輸系統，物件為基礎之視訊摘要可以用來對物體所發生之重要事件發出警訊。此外，視訊內容摘要也可以幫助我們從監控視訊影像的資料庫中取得與管理重要的內容。在這篇論文中，我們提出了一個以物件為基礎之監控視訊之影像內容追蹤與摘要的系統。首先，我們使用以背景作為參考之動態物件切割演算法把運動物體從背景中分割出來。接著，我們使用一個簡單但有效的動態物件追蹤演算法得到運動物件的軌跡與特徵。最後，我們設計了一個藉由選取代表著重要影像內容的動態物體來產生影像內容摘要的演算法。我們用此演算法來產生以物件為基礎之影像內容摘要。

我們藉由展示出我們所選取到具有代表性意義與事件的物件來證明我們提出的系統的成效。根據所提出的系統架構，我們實作了一個能夠線上發出警訊之即時影像監控系統。在這個系統中，我們可以藉由所產生的內容摘要即時地對重要的事件發出警訊並且同時對監控影像中的物體做即時的追蹤。我們測試了一些不同類型的監控視訊影像的片段，而實驗結果證明，我們的系統在物件內容之追蹤與在重要物件內容之摘要都得到滿意的結果。

Video abstractions represent the video contents concisely by extracting the important semantics in the video. Important events and semantic meanings of moving objects in the surveillance videos can be represented using the object-based video abstractions. Therefore, the object-based video abstractions can be used to alarm important events in the surveillance and monitoring systems like the intelligent transportation system (ITS). Besides, the video abstractions can also help retrieving and managing important contents from the surveillance video database. In this thesis, we propose an object-based video tracking and abstraction system. First, we use a background-registration segmentation algorithm to segment the moving objects. Then, a simple but effective tracking algorithm is introduced to extract the object trajectories and object features. Finally, an abstraction algorithm is applied to generate object-based abstraction by selecting key objects which contain important semantic contents.

We will reveal the performance of the tracking and abstraction algorithms by showing the results of extracting objects with representative features and events such as object appearance/disappearance, objects occlusion/split and changes in motion. We will present a real-time tracking system with on-line alarming to demonstrate the implemented system. In this system, important object events will be alarmed using the generated object-based abstraction on-line while the surveillance video is being tracked in real-time. We test our system with several surveillance video sequences and the experimental results prove that we can get satisfactory results in both the object-based tracking and abstraction.