Influence of different ground truth hypotheses on the quality of Bayesian networks for maneuver detection and prediction of driving behavior

Semi or fully automated driving requires anticipating the driving behavior of traffic participants to increase safety, comfort and a cooperative driving experience. For the task of prediction machine learning techniques are often applied to learn driving behavior from measured data. Due to the very subjective nature of predicting driving maneuvers, in this paper, the significance of the effort and time spent in the right kind of labeling for time series data and its direct relevance to the performance of the trained models is shown.