Driver cognitive fatigue can significantly affect driving and may lead to fatal accidents. In this regard, automatic detection of underload driver cognitive fatigue based on upper body posture dynamics is studied in this paper, where a semi-supervised approach is developed to identify the cognitive fatigue patterns of driver posture. Initially, an unsupervised Gaussian Mixture Model (GMM) clustering is applied to the acceleration data representing the driver's head, neck, and sternum obtained in a simulated driving through a motion capture suit. This provides the optimum clusters of the most-similar and correlated time-series data of driver upper posture. Then, an automatic labelling algorithm is developed that mines the maximal value and the standard deviation of each GMM cluster and assigns a symbol according to the discrepancy in postural behaviour. Finally, novel machine learning supervised classifiers, including Gaussian Support Vector Machines, and Bootstrap-Aggregating base