Prediction of Road User Behavior

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
The workspace of advanced driver assistance systems is getting more complicated as they start targeting more complex environments. For ensuring the safety performance of these systems, recognizing the surrounding traffic behavior is not sufficient: the prediction of the behavior of other road users is becoming essential. The main objective of this study is to design the system that can predict the trajectories of road users within the 3-second prediction horizon. Furthermore, the objective also includes making the toolchain for data processing, training, and evaluation of the trajectory prediction system. A deep machine learning method was proposed for the problem of trajectory prediction. The solution was comprehensively studied on the naturalistic driving data that was available at TNO. The proposed solution concept is proven as an efficient method for predicting the road user trajectories. Although there is the clear evidence that the model should be trained with more data, the study shows that the model has learned important properties, and it is consistent in all the performance measures. The models are evaluated using five metrics: position accuracy; lane change precision, recall, F1 score, and end lane accuracy. The proposed method sufficiently outperformed linear extrapolation in the lane change prediction and showed similar to linear extrapolation position accuracy. Moreover, the method correctly models prediction variance that can be viewed as the quality estimation of the prediction a priori.
Faculteit der Sociale Wetenschappen