本論文目的在於建立一套基於智慧型手機的即時情緒辨識系統，藉由分析光體積變化描計圖(Photoplethysmogram；PPG)來辨識情緒。研究中辨識的情緒一共五種：平常狀態(一般不受刺激的情緒)、快樂、壓力、悲傷、生氣。受測者為十位男性與十位女性，使用的情緒刺激源為觀賞一段二至四分鐘的影片。研究前期先於Matlab驗證出較好的演算法，後期將前期研究的結果轉至Android平台以實現即時辨識的系統。本即時辨識系統是在Android平台上，藉由藍芽傳輸實現無線光體積變化描計圖的情緒辨識系統，Android平台的功能可分為訊號擷取、P點偵測、特徵計算、支持向量機(SVM)分類及辨識結果。本研究特別著重於探討不同情緒向度在情緒辨識的效果，探討的情緒向度為情緒向性（valence）和激發水準（arousal）。實驗結果顯示情緒向性可得到較好的分類，當使用基因演算法(GA)挑選最佳的特徵組合，並以SVM做為分類器，在將情緒以情緒向性分類成兩大類的情緒狀態，使用留一交叉驗證可得到91%的準確度。另外當使用二階段式分類法分辨五種情緒時，將兩個情緒向度的結果加入原來的特徵集中，正確率可從原先的62%提升至80%，大幅提升分辨多種情緒的正確率。 This paper aims to develop a real-time emotional states recognition system on smartphones, which analyzes Photoplethysmogram (PPG) for emotion recognition. There are five emotion states used in the study, including normal, happy, sad, stress, and anger. The participants include 10 male and 10 female students who watched video programs associated with different emotions of two to four minutes in length to stimulate distinct emotions.In the early stage of the research, Matlab was used to construct a superior algorithm. In the later stage of the research, the algorithm developed in the previous stage was transferred to the Android platform to achieve a real-time recognition system. The real-time emotion recognition system is developed on the Android platform. PPG signal is transmitted through Bluetooth wireless transmission. The functional blocks of the Android platform include signal capture, P-point detection, feature calculation, classification and result display on the screen.A total of 20 subject with equal number of male and female participated in the study. The PPG signal was segmented into 20-sec segment for emotional analysis. The study is focused on exploring the effects of different emotional dimension on recognition accuracy. Two emotional dimensions, valence and arousal, were considered in the study. In this study, the valence showed better discrimination power as genetic algorithms (GA) being applied to find the best feature set. And a high accuracy of 91% was achieved by Support Vector Machine (SVM) classifier with leave one person out Validation. For multiple emotion states recognition, adding the results of dimensional discrimination based on valence and arousal into the original feature set, the recognition rate can be boosted from 62% to 80%.