Abstract:
In recent times an increasing number of drivers are injured or killed in road accidents. The majority of these accidents are caused by drunk driving, driver fatigue, distractions from outside the vehicle, and so forth. However, if a vehicle includes a passenger sitting next to the driver, the probability of a car accident drops significantly. This is because the passenger can alert the driver when the driver is distracted or the driver's condition is not suitable for driving, such as when he or she shows signs of fatigue. Coinciding with the rapid development of computer vision technologies, researchers have tried to find ways to assist drivers if there is no passengers present in a car or if the passenger is unaware of risk factors. Thus, vision-based driver assistance systems have attracted a great deal of attention. Estimating a driver's head pose is an important task for driver assistance systems because it can provide information about where a driver is looking, thus, providing useful cues on the status of the driver (i.e. paying proper attention, fatigue, and so forth). This work proposes a system for estimating head pose using monocular images and includes a novel use of back-projection. The system can work on a single image to provide an estimation of a driver's head pose at a particular time stamp, as well as on an image sequence to support an analysis of a driver's status. Within the proposed system, two previously known pose estimation approaches are implemented for comparison - PnP and POSIT. Also an approach is introduced to provide ground truth reference data by using a mannequin model. Experimental results demonstrate that the proposed system is able to provide relatively accurate estimations of yaw, pitch, and roll angle. They also show that PnP provides better estimation than POSIT within the proposed system.