手勢辨識的演算法在效率與學習資料兩點會對應用時造成使用上的不便利,諸如需要事前定義好追蹤目標,需使用大量影像作為索引,限制了需使用的記憶體空間;而這些缺陷也違背了人機互動(HCI、Human-Computer Interaction)系統講求自然操作的原則。因此,如何在不定義追蹤目標,又只使用手勢本身的邏輯結構來辨識手勢,除是重要課題外,也是日後應用於其他系統中必須克服的環節。 本文提出階層式手勢模型(HHM、Hierarchy Hand Model)作為擷取特徵的方法,一來利用手勢特有的邏輯結構當成辨認的要件,除去需事先定義追蹤目標的缺點,二來由模型產生的特徵參數,可直接運算出手勢的意義,取代演算法中複雜且龐大的模型或影像資料庫外,也避免因使用學習系統造成的結果不正確性與冗長的學習過程。在追蹤演算法上使用Mean Shift與Particle Filter混合的追蹤演算法,並於前置處理中,用ROI-MAP來縮小偵測區域以及對影像最好事前處理,減少誤判的發生。辨識方面,本文將手勢分成靜態手勢與連續手勢,將手勢的連續變換動作組合成一狀態序,以供後續系統開放使用。
The algorithm of hand gestures tracking and recognition using in application has two problems that effect the system. One problem is the constrain before using system, such as define of target color. The other problem is the necessity of huge image data of training learning system, which will constrain the minimum memory size. However, the principle of Human-Computer Interaction is naturally to be used, the above problems will enable the un-nature use. Hierarchy Hand Model(HHM) is a feature that can solve above problems by using hand logic. To solve the first constrain, HHM defines hand shape to use hand logic, then it is not necessary to define the target color. For the second constrain, HHM using hand logic to detect hand parameter, so that the recognition can be used by the physical models. The tracking algorithm using HHM that has a problem of performance. However, Kernel-Based object tracking has good performance and outcome when mix Mean-Shift and Particle-Filter algorithm. In preprocess, using ROI-MAP to reduce tracking region. In recognition, divide hand gestures for static gestues and continuous gestures that is the signal can easily be used in application system.