肢體語言辨識（Gesture Recognition）為人機互動（Human Computer Interaction）運用中一項重要的技術，其中視角無視辨識（View-Independent Recognition）為機器視覺辨識的難題。為了賦予機器學習模型（Machine Learning model）辨識事物於不同視角的能力，時間資訊的運用是一道線索。然而多數的機器學習模型本質為辨識模型（discriminative model），運算複雜度的問題使其困難於運用時間資訊，並被認為欠缺對輸入訓練資料的歸納性（generalization）與藉由過去經驗幫助新事物學習，增進學習（incremental learning）的能力。 分級時序記憶（Hierarchical Temporal Memory）為近年新發展的機器學習模型。根據人類大腦皮質的運算假說：記憶預測架構，建構非辨識模型（non-discriminative model）。分級時序記憶利用時間資訊行使非監督式學習，使機器學習模型具備歸納訓練資料與增進學習的能力，同時達到可信賴的辨識結果。本論文使用電腦視覺演算法與分級時序記憶實作兩個手勢辨識問題，於視角變動的連續影像的單張辨識中（snap shot）分別得到辨識正確率91%與84%的辨識結果。 Gesture Recognition is importance in designing efficient Human Computer Interaction (HCI) applications and View-Independent Recognition is one of a difficult computer vision gesture recognition problem. Temporal information is a clue to provide the ability to recognize object in variant phase for Machine Learning model. However, most of the Machine Learning Model is discriminative model. It has computational complexity problem for using temporal information and proves inadequate at the ability of training data generalization and incremental learning essentially. Hierarchical Temporal Memory is a novel Machine Learning model studying in recent years. According to the memory prediction framework hypothesis of brain new cortex, Hierarchical Temporal Memory builds a non-discriminative model using temporal information to do unsupervised learning. Try to achieve training data generalization and incremental learning ability without losing recognition reliability. Combining computer vision image process algorithm and Hierarchical Temporal Memory Machine Learning model, a hand gesture recognition system was built in this paper. Two continuous view-point change recognition problems was tested, the continuous image sequence snap shot recognition accuracy results were 91% and 84% respectively.
