Electronic Thesis and Dissertation Repository

Thesis Format



Master of Science


Computer Science


Dr. Steven Beauchemin


In many advanced driver assistance systems (ADAS) applications, it is essential to figure out where gaze of driver locates in image area of stereoscopic vision system. This problem, which involves a cross calibration between the stereo vision system and eye tracker, is a challenging task since the two systems are not consistent in modality and do not share a common image area. The eye tracker system provides a 3D gaze vector which describes the direction of driver’s 3D line of gaze, while the stereoscopic vision system provides a depth image frame. In this thesis, this crosscalibration was performed with a closed-form solution that employs an efficient, linear time Perspective-n-Point algorithm. The main contribution of the present thesis is reformulation of this cross-calibration problem in a way that we would be able to employ PnP algorithm for providing a closed-form solution. The calibration process maps the 3D driver gaze vector into the surrounding outdoor environment. Moreover, the robustness of the algorithm with respect to noise is investigated on a set of synthetic data as well as in a lab-environment place. Keywords: Cross-calibration, perspective-n-point, driver

Summary for Lay Audience

Advanced driver assistance systems (ADASs) have already been implemented in vehicles. Such examples are cruise control, warning systems for lane departure and automatic parking that have been recently introduced.

ADASs are supposed to provide a safe driving experience, but fatalities involving these systems increased during recent years. It is believed that if their full potential be realized, ADASs would have an annual benefit of around $800 billion by 2050 via mitigation in traffic congestion, energy consumption and road collisions. This goal cannot be achieved unless knowledge in ADASs is advanced to a greater extent.

ADASs are focused on perception of environments around the vehicles for assisting the driver, while little attention is given to the perception of driver behavior. It is believed that driver behavior can significantly improve ADASs as 95 percent of vehicle collision are due to human error. For detection of visual driver distraction, it is important to find out where the driver is looking. This is the main subject of current thesis.