Eye movement feature extraction for driver vigilance classification
Date of Issue2015
School of Electrical and Electronic Engineering
Along with the advanced and mature development of eye tracking technology, eye movement tracking shows great potential in the construction of driver assistance system, which may provide services for driver at early stage of low vigilance. However, the raw data from eye tracker is time dependent and cannot be directly utilized for machine learning. Also, the existing feature extraction software Tobbi Studio cannot be customized based on needs and is not real time. Hence, a new eye movement feature extraction tool needs to be developed for the convenience of future research. In this report, the development of an eye movement feature extraction tool based on Velocity and Dispersion Threshold Identification (I-VDT) algorithm was introduced. This tool may extract fixations, saccades and smooth pursuit these three fundamental eye movements with satisfactory accuracy. Especially for the extraction of fixations and saccades, accuracy was generally above 95%. Through experiments, recommendations were given on parameter setting. 30º/s (º/s means deg/sec in this report) is suggested to be the velocity threshold that separates fixations and saccades. Also, the performance of the algorithm is not sensitive to the minimum fixation duration threshold in the range of 60 seconds to 100 seconds. In the second part of the project, statistical analysis was conducted on twelve eye movement feature metrics based the data collected from Vigilance Decrement Experiment, which is run by NTU HMI Lab. Analysis of variance (ANOVA) test was employed to check if there exist significant differences among three stages within one driving period. Both mean value and variance (standard deviation) of fixation duration, fixation centroid location, saccade peak velocity/duration, saccade velocity and saccade duration presents differ among stages with 95% confidence level for more than half of the subjects. Hence, these metrics are potential to be utilized in machine learning.
DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
Final Year Project (FYP)
Nanyang Technological University