A Robust 3D Face Model-based Head Pose Tracking Approach for Driver State Estimation

摘要

Driver distraction and inattention are the prominent causes of automotive collisions. To address these problems in the driver assistance system, in this paper we present a completely automatic robust three-dimensional (3D) head pose tracking method in vehicle environments, by which a new sensing approach to infer a driver's focus of attention is available. After the automatic initialization of finding the face region and estimating head pose is finished in the first frame with the approximately frontal face, all energy of our method is devoted into the processing procedure in the successive segments with two continuous frames. First, the KLT algorithm is used to detect good points in the face region in the front frame and track them across to the latter frame. In this way, the matched feature points in each successive segments are readily computed, which is used as input in the model-based bundle adjustment framework to recover head pose of the 3D face model and obtain the head 6-DOF (including 3 translations and 3 rotations) information in cockpit. Extensive experiments are carried out at actual driving environment to demonstrate that our work has the capable of dealing with partial occlusion, large motion (include rotation).