Driver Verification Using Eye Movements and Blinking
Author(s): Zandi, Quddus, Prest, Comeau
Slidedeck Presentation Only (no paper submitted):
Abstract:
With the advance of technologies used in vehicles and from road safety point of view, driver authentication would be required in various contexts including verifying the identity of the person operating in-vehicle safety devices (e.g. alcohol interlocks) or the driver in a connected vehicle in case of accidents or traffic violations. While various types of biometric modalities are currently available, many of these technologies either are not robust against circumvention/spoofing (e.g. facial recognition or fingerprint) or interfere with the driver activity; hence, they are not appropriate in case of a vehicle in motion. The goal of this study was to investigate the characteristics of eye movements and blinking for verification of drivers. Ultimately, this research would lead to development of non-intrusive techniques for practical and reliable driver biometric verification. Eye tracking data was acquired from 30 volunteers (age 40.13±9.69 years; 8 females) using an advanced infra-red-based system with two cameras at the internal Research Laboratory of Alcohol Countermeasure Systems Corp., Toronto, Canada. Each subject participated in two separate sessions using a driving simulator, following informed written consent: a 10-min morning control driving (CD) session and a 30-min monotonous driving (MD) session in the afternoon. The multidimensional eye tracking recording from every driving session of each participant was then segmented into 10-sec timing windows with 5-sec overlap, and 28 features were extracted from each window. The features were extracted from four main categories: general gaze, fixation, saccade, and blinking. The resulting feature vectors were then presented to two non-linear classifiers: a support vector machine (SVM) with radial basis function kernel and a gradient boosted tree (GBT) with maximum tree depth of 5. The final prediction was done based on the fusion of scores from the two classifiers. We evaluated the performance of the eye-tracking-based features in verification of each subject using the SVM and GBT classifiers for CD, MD and mixed (CD and MD) sessions. The evaluation was conducted by splitting data randomly to 90% training and 10% testing, while the whole procedure was repeated 10 times. According to the results based on score-level fusion, at the false acceptance rate of 1%, the false rejection rate of 5.13%, 8.06%, and 18.77% was achieved for CD, MD, and mixed sessions, respectively. The equal error rate for these three scenarios was, respectively, 2.43%, 3.75%, and 6.67%. Results clearly demonstrate that driver verification using eye tracking features can be implemented with high accuracy. Overall, the score-level fusion improves the accuracy, although accuracy decreases in case of mixed sessions. The ability to verify in mixed sessions (CD and MD) is a highly desirable feature for any practical biometric systems. The proposed biometric verification system is suitable for practical implementation using simple features that may be extracted very efficiently in real-time from commercially available eye trackers. This verification scheme can be easily implemented on embedded platforms. More investigations under real driving conditions with a larger number of subjects are planned for the future work.