Methods for robust driver pose estimation.

摘要

This thesis presents methods for monitoring a driver's pose while he or she operates a vehicle. The driver pose is found by analyzing a video of a driver taken from a dash-board mounted video camera. Image frames are extracted from the video and used for analysis. The investigated approaches can be broadly divided into the following categories: (a) Pose estimation methods, which are used to determine the driver pose for each frame of the video. (b) Combination of classifiers, which use independent pose estimation algorithms to determine the driver pose for each frame of the video. These classifiers also have the option of rejecting video frames which cannot confidently be classified.; In addition, a method for estimating driver motion is presented. The principle contributions of this thesis are a new innovative pose estimation approach, a motion estimation algorithm, and three combination of classifiers which have a rejection mechanism. Additionally, this thesis investigates the analysis of the driver under two sets of pose classification guidelines.; Simulation results show the performance of the system with respect to six driver video databases.