Attention estimation from the integrated IMU and eye-tracking perception of smart glasses

• Main Authors: CHEN, PIN-CHIH, 陳品志
• Other Authors: CHEN, TZU-CHIANG
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/50219533067558386819

碩士 === 國立中正大學 === 電機工程研究所 === 104 === Wearable mobile learning becomes an inevitable trend in the future. With the matured wearable technologies, related applications have boosted in the market, but lacked for a killer application that can lead and spread wearable technologies. In another way, attention ability is essential to effective learning. By estimating and recording attention capability help learners to improve and review their learning process. In the light of this, attention-estimating system is constructed form IMU information, and eye-tracking perception in our study. In this study, we combine eye-tracking perception with IMU information to realize the Attention-estimating system based on Android smart glasses. First, we divided the eye-tracking into eyeball detection and eye corner detection. Minimum Average Gray Value was used to find the general eyeball contour. Moreover, considered all kinds of light changes, Otsu algorithm also was used to detect the contour of eye corners. Then, we can master the eye movement by determining the sight and projection point. The result of the average error angle of view is 1.82 degree. In the attention recognition, the system divided steps into five part: data acquisition, feature extraction, feature selection, classify and voting mechanism in this study. In order to extract the eye movement and IMU information by recording experiment of attention and non-attention in seven different scenarios; In the feature extraction, we extracted forty one kinds of features based on feature characteristics, and selected the suitable feature by four different kinds of feature selections; Finally, we optimized the parameter of SVM by Genetic algorithm, and validated the reliability by K-fold algorithm, however, we also adapted the voting mechanism to estimate the result of attention level. The classified accuracy of attention and non-attention are reach up to 86.10% under seven scenarios. The three and two categories of Continuous Performance Test experiment result are 81.12% and 83.44%, respectively. Eventually our result as compared with reference papers that indicated feasibility and reliability.