以平行多維數動態時間校正辨識連環手勢 的方法

Abstract

動態時間校正（Dynamic Time Warping）是個針對不同時間長度的兩種訊號歸類的務實方法。動態時間校正所提供的對應點（Alignment）能夠有效的把兩邊的訊號全部對應起來,重而計算出兩者的相似值。這一點對於手勢辨識是相當有用的。然而動態時間校正本身並不完善至足以應對即時連環手勢辨識。此論文針對即時連環手勢辨識呈現兩種演算法。第一種演算法為平行動態時間校正。它能在數據不充足的情況下，依然能準確的辨識出正確的手勢。第二種演算法為自動計算門檻值。一般上此門檻值做為資料分析過濾用途。該演算法成功簡化原本手動調整門檻值所帶來不必要的麻煩。 這兩種演算法使即時連環手勢辨識更能發揮其效用。

Dynamic Time Warping is a handy solution in matching time series sequences of different lengths. The alignments provided by Dynamic Time Warping are meaningful in matching both time sequences, proving its usefulness in hand gesture recognition field. However it has flaws where it is not robust enough in imperfect situations like inadequate amount of input data. In this thesis, two algorithms are proposed, where the first one is Parallel Dynamic Time Warping (PDTW) as replacement to Dynamic Time Warping to provide a more robust way to recognize continuous hand gestures; and the second algorithm is Automatic Threshold Determination, which automatically defines threshold values. These threshold values used to be manually calculated by user and the calculation process can be rather fussy. Both algorithms smoothen the process of multi-dimensional hand gesture recognition.