以生理訊號分析系統即時評估音樂環境之使用者情感反應

Abstract

情緒辨識技術結合手機等可攜式裝置，可提供人與人之間更完整的溝通資訊，同時增加人機互動的豐富性。利用相關的生理資訊可以建立單人的即時情緒辨識系統。本研究以音樂來改變使用者的情緒，使他們產生逐漸放鬆、愉快(正向)及不愉快(負向)三種情緒反應，同時蒐集使用者的肌電圖、呼吸、脈搏及表面皮膚導電度。蒐集的生理訊號經過濾波、切割、校正與正規化後可以得到相關特徵。同時，利用生理特徵配合分類器，找出對於情緒判定最有用的生理訊號。 離線結果部分，單人放鬆 vs. 強烈反應的辨識正確率達95.61%，正向 vs. 負向辨識正確率達91.69%。另外，利用單人多次實驗結果觀察使用者皮膚導電度在不同情緒狀態下的反應趨勢，結果得到在放鬆狀態，導電度逐漸下降，而強烈反應下，導電度會有上升情形，符合文獻結果。 即時結果部分，單人即時放鬆 vs. 強烈反應的辨識正確率達94.69%，正向vs. 負向辨識正確率達81.00%。 本研究最後對於利用生理訊號即時判斷單人情緒所遇到的種種問題加以探討，並對未來建立單人即時情緒辨識系統所需的努力方向提出說明。

Integration of emotion recognition and portable devices such as cell phone could provide more completed information for people communication and better human-computer interaction. A real-time emotion recognition system for individuals could be implemented with related bio-information. In this research, specific music is chosen to elicit the user’s emotions (relaxed, positive and negative). The physiological signals were acquired through four biosensors: electromyogram, skin conductance, respiration and pulse. Physiological features are acquired by features extraction methods such as filtering, segmentation, calibration and normalization. At the same time, physiological features are classified using pattern recognition techniques. The accuracy of off-line analysis achieved 95.61% and 91.69% on recognition of “relaxed vs. excited” and “positive vs. negative”, respectively. Besides, our results show the tendency of user’s skin conductance responses matches other research results. Furthermore, the accuracy of real-time analysis are 94.69% and 81.00% on recognition of “relaxed vs. excited” and “positive vs. negative”, respectively. Finally, the limitations of real-time emotion recognition for individual are listed and will be solved in the future; there are still some works need to be optimized for implementation of a real-time emotion recognition system for individual.