Abstract
This thesis proposes an intelligent vehicle system to identify the driver improper driving behavior using the so-called “first-person vision” (FPV) technology. Different from conventional computer vision, FPV is based on the person wearing a goggles camera representing the subject vision. There are two technologies proposed in this thesis: vehicle exterior/interior view detection and driver viewing angle estimation. For the first method, we use “bag of words” image classification approach by applying FAST+BRIEF feature descriptor in the dataset collected in advance. Then, we establish the first-person vision “vocabulary dictionary”, encoding an input image into a feature vector. Finally, we apply SVM classifier to detect whether the input image is from the inside of a vehicle, and further identifying the driver current attention. For the second method, we install an extra vehicle-mounted camera to record the image in front of the vehicle and to deem it as world coordinate reference for viewing angle estimation. Then, we find the relationship between the world and first-person-vision camera coordinate. Finally, we may further estimate the viewing direction of driver.