在人機介面之互動上,透過使用者的情緒表現來瞭解其感受回饋是相當重要的。本研究之目的即在於擷取及分析生理訊號並進行情緒辨識分類。研究中透過IAPS(International Affective Picture System)誘發多位受試者強烈正面、微弱正面、強烈負面、微弱負面等四類情緒之產生,另也透過事先挑選之影片誘發多位及單一使用者好笑、愉悅、噁心、害怕四種情緒之產生,並同時利用生理訊號感測器量測肌電圖、心電圖、呼吸訊號及脈搏訊號等生理訊號。收集之生理訊號經過正規化、分析處理及特徵值擷取後,將33個特徵參數透過KNN (k-nearest neighbor algorithm)分類器進行分類,以達到辨識情緒的目的。 結果顯示,利用IAPS圖片系統、利用影片刺激多位使用者,及利用影片刺激單一位使用者,以KNN為演算法將所得之33個特徵值進行情緒分類,其辨識率為90.87%、95.32%、96.58 %。再透過資訊增益(information gain)選出前10個重要的特徵值,分類辨識率則為91.45%、96.1%、97.61%。 結果中也發現由於情緒表現於生理訊號上較具有個體特異性,因此若是利用非本身之情緒資料來判斷其情緒類別,辨識能力僅約31.93 %。但若是利用其受試者本身之情緒資料來進行分類,其辨識率則有95.47 %。因此顯示個人差異性對於辨識率具有極大的影響,需要持續增加使用者個數以降低個人差異性。 本研究最後對於研究使用者情緒辨識時所碰到的難題加以討論,並對未來進行即時使用者情緒辨識系統的研究提供方向與策略。
It is important to understand the user’s feeling and feedback by user’s emotional expression in human-computer interaction. The aim of this study is to develop a affective response recognition system by bio-signals measurement, feature extraction and classification. The IAPS (International Affective Picture System) is adopted to elicit user’s affective responses included high valence high arousal, high valence low arousal, low valence high arousal, and low valence low arousal. Moreover, the prepared video clips are also used to elicit multi-user’s and single user’s affective responses included laughing, pleasure, disgust, and fear. The user’s physiological signals, EMG, ECG, blood pulse, and respiration signal, would be measured and recorded simultaneous. By normalization, signal post-processing and feature extraction, biophysical signals would be classified by KNN classifier to indicate the corresponding affective response. The results show that the accuracy of using IAPS, Video to elicit multi-users’ affective response and using Video to elicit single user’s affective response are 90.87%, 95.32%, and 96.58 %. If data with only top-10 important features which obtained by evaluating the information gain is used to classify, the accuracy would become 91.45%, 96.1%, and 97.61%。 If the data of specific one user is used as the testing dataset and other is used as the training dataset, the accuracy would drop off only 31.93%. The main reason is the number of user is not enough, and the differences between each individual are obviously. Besides, if we separate the data of specific one user by n-fold cross-validation as the training dataset and testing dataset, the accuracy would maintain higher about 95.47%. Therefore, in order to overcome it, more experiments and more biophysical signals are necessary due to the difference between each individual would affect the accuracy of recognition seriously. At the last of this thesis we have discussed about the tough question of studying user’s affective response recognition system and provided the suggestions and strategies to a real-time user’s affective response recognition system.