Optical flow preprocessing for pose classification and transition recognition using class-specific principle component analysis

This paper introduces a technique for achieving a high degree of accuracy in driver pose estimation and pose transition identification using pose-specific principle components developed by Belhumeur, Hespanha and Kriegman (1997) in a noisy video sequence. The proposed method uses the estimated optical flow in a bandpass filtered image series to establish a spatially stable window that defines the pixel domain for further processing and principle component extraction. Optical flow tracking is achieved via spatial correlation using a likelihood measure that accounts for similarity between pixel values and brightness distribution in sequential video frames.