Vehicle lateral position prediction: A small step towards a comprehensive risk assessment system

The current paper explores the forecasting capabilities of a Multi-Layered Perceptron employing a sliding input window (MLP-sw) with respect to lane change and overtaking maneuvers. The performance of the network was assessed in terms of prediction accuracy at different prediction horizons. Test results using driving simulator data showed that the maximal prediction error at a 1s horizon is well within tolerable range, and that prediction horizon of 2s or more can potentially be brought into an acceptable range with output smoothing and other means. One reason that prediction accuracy decreases as prediction horizon increases is that when a larger amount of maneuvers are involved, it becomes increasingly difficult for the system to disambiguate among multiple dependencies. Some potential ways of improving prediction accuracies, as well as the inherent challenges of forecasting vehicle positions, are discussed.