A System Using Photoplethysmography for Emotion Recognition Based on Smartphone

• Main Authors: Po-Kai Huang, 黃柏凱
• Other Authors: Sung-Nien Yu
• Format: Others
• Language: zh-TW
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/42158440955373104640

碩士 === 國立中正大學 === 電機工程研究所 === 103 === This thesis presents a emotion recognition system based on analyzing photoplethysmography(PPG) on smartphones.This system focuses on the effect and accuracy of the real-time recognition system,thus short data(20 seconds) are used.In this study,we firstly classified,the signalinto three kinds of emotions (positive、 negative、neutral).However, in this condition,were of unequal sample numbers in each class.Therefore, we secondly explored the method and effect of and adjusting the class samples to be equal number. Finally,we explored the performance of the system in recogniting six emotions,namely normal state,happy,sad, stress,anger,disgust.The participants consistd of 10 male and 10 female students who watched video programs associated with different emotion of two to four minutes in length to stimulate distinct emotions. In this study,we used genetic algorithms (GA) to select an effective combination of features to enhance the recognition rate.The support vector machine (SVM) was used as classifier and leave-one-out and all-train-all-test methods were employ to validate the system. By using leave-one-out cross validation and genetically algorithms to select the most effective features,an accuracy of 78.3% was achieved by using the original PPG features to classify the signal into three emotions.After matching the number of samples in each class,the accuracy in classifying emotions was 62.5%.The recognition rates of the unmatched samples favered the negative emotion while the matched samples produced more consistant rates across different emotion categories.The accuracy of recogniting six emotions was 55%. The study also found a very important indicator in the waveform of PPG called the Dicrotic Notch,which is closely related with the function of human’s autonomic nervous system,thus we recruited new feature associated with the indicator in the study.After adding dicrotic notch features,the accuracy in classifying three emotions raised to 84.4% and 67.5% with the unmatched and the matched samples,respectively.The accuracy of recogniting six emotions was 60.8%. In addition,we also considered to include difference features to exclude individual difference,which is practically achieved by subtracting baseline feature values from the active feature values.After including the dicrotic notch and difference features in the feature set,the accuracies in classifying were raised to 92.5% and 80.8% with the unmatched and matched samples,respectively.The accuracy of recogniting six emotions was 65%.Therefore,after feature selection,with the added dicrotic notch features and difference features,the accuracy can be effectively improved with the GA feature selector. Finally,we implemented the sensing and processing system with wireless bluetooth transmission of the PPG on an Android platform. The users interface present the result of the system classification of signal into six emotions,namely, happy,sad,stress,anger,disgust and normal.When considering baseline measurement,the average accuracy was 62.5%.When not considering baseline measurement,the average accuracy was 67.5%.The result demonstrated the high-capability and real-time recognition feasibility of the proposed emotion recognition system.