結合動作感測器之角色動作合成技術

Abstract

近年來，隨著動作感測器的逐漸普及，以體感控制為基礎的遊戲與人機介面成為數 位內容相關領域中最受注目的新發展。然而，受限於平價儀器的準確度與感應器的數 目，這類的介面方式往往無法反應使用者快速與多變的動作，僅能以簡單動作指令來控 制操作。 本研究計畫，針對此問題，提出以分析比對感測器資料與大量擷取之人物動作資料 庫，藉由結合人物動作辨識與動作合成技術，能以少量的動作感測資料推估出使用者對 應的動作變化。本計畫預計以三年的時間完成。第一年將利用電腦學習技術，著重動作 分析與辨識；第二年將動作辨識技術與以範例資料庫為基礎的動作合成技術結合，完成 更擬真的體感介面。第三年將更進一步發展體感資料與人物動作對應技術，以合成與控 制不在資料庫中之動作。 本計畫所發展的技術，能以普及的平價動作感測器即時產生平順擬真的人物動畫與 體感動作控制，將可應用於擬真體感遊戲或是更進一步的體感人機介面。

Recently, motion-sensing-based games and interfaces become the most attractive topics in related fields of digital contents. However, due to the accuracy and amount limitations of popularly used motion sensors, users can only perform simple and restricted motion commands with these kinds of interfaces. In this project, we propose analyzing and comparing motion sensing data with human motion capture database. We plan to combine motion recognition and synthesis techniques and estimate high-resolution character motion from sparse motion sensing data. In the first year, machine learning methods will be utilized for motion analysis and recognition. In the second year, motion synthesis techniques in computer animation will then be combined with the recognized action commands. A more detailed character motion can be evaluated and synthesized. In the third year, we plan to develop a more advanced character motion mapping technique to deal with motions not included in the trained categories. With the proposed methods, more smooth, realistic and high-resolution character animation and interactions can be realized. These techniques can be applied to realistic motion-sensing games or advanced human computer interaction.