基於人體動作分析之真實影片卡通特效化

陳澤智
The principal objective of this paper is to build a novel video tooning technology, and it contains two principal components: hierarchical video segmentation and video tooning with special effects. In the first part of the system, we through the human motion build a video structure, which includes Entry shot, Action shot, and Exit shot respectively. The Action shot is the core of the video, which is divided into Gimmickry period, Incubate period, and Shooting period.

Through the video structure, the system can efficiently compute some parameters, and these parameters interact with special effects, which we call special effects parameters. In the second part, we use mean shift segmentation to segment the foreground object. And we use the periods of video structure to compute these special effects parameters separately and through these parameters, they interact with special effects. Furthermore, there are three interactions with the special effects, which include special effect interacting with environment, with other special effects and with attacker. The contribution of this paper includes three major components: (1) The system analyzes the human motion to build a video structure which cuts video by hierarchical video segmentation, dividing two layers into shots and periods. (2) The system can recognize human motion automatically and combine with the special effects in order to design the dynamic special effects. (3) Through the combination of the special effects and parameters, the system makes the three interactions— with attacker, environment, and other special effects. The system gets these interactions with special effects that make our results more interesting and vivid than traditional video tooning.

本論文旨在建立一個視訊卡通特效化的系統，本系統分為兩大部分，分別為hierarchical video segmentation與video tooning with special effects。在系統的第一部分中，首先利用人體動作定義一個video structure，包含Entry shot, Action shot, and Exit shot respectively，其中Action shot為影片的核心，因此會再被分為Gimmickry period, Incubate period, and Shooting period。透過video structure幫助系統有效率的找出與特效互動的參數，我們稱之為特效參數。在系統的第二部分，我們利用mean shift segmentation對前景做卡通化。並且利用video structure切出的periods，分別求其所屬的特效參數，透過特效參數達到與特效的互動。此外特效的互動關係有三種，除了特效與人動作的互動之外也包含特效與環境的互動以及特效與特效的互動，系統透過卡通與特效的結合，使我們的卡通結果能有更多元的面貌呈現。本篇論文的主要貢獻包含三個部分：(1)系統分析人的動作建立一個video structure，將影片作階層式切割，分成shots與periods共兩層。(2)系統自動辨識人的動作，並將之結合特效，進而產生出動態的特效。(3)透過特效與特效參數的結合，特效擁有三種互動—與人的動作、環境以及其他特效。系統透過特效與互動的結合，使我們的結果比起傳統的真實影片卡通化更加生動有趣。