Lane-changing prediction in highway: Comparing empirically rule-based model MOBIL and a naive Bayes algorithm

Many models and algorithms allow predicting driver lane-changing intents in highway. Generally speaking, the challenge consists of inferring lane-changing maneuvers from speed difference and spacing with the surrounding vehicles on current and intended lanes. In this contribution, we empirically compare two approaches: the MOBIL model and the naive Bayes algorithm. The model MOBIL is a well-established rule-based approach, while naive Bayes is a data-based classifier by machine learning. The analysis is done using naturalistic trajectories of two-lane German highways (HighD project). We identify characteristic relationships between the spacing and speed difference variables and the intent to keep lane, overtake, or fold-down. It turns out that the mechanisms initiating fold-down and overtaking are different, requiring analysing the maneuvers separately. The fold-down maneuver is a more complex process involving more surrounding vehicles in interaction. False-positive and true-negative prediction errors of lane-changing and lane-keeping intents are computed using cross-validation. The quality of prediction with the rule-based model is satisfying for overtaking and limited for the fold-down maneuver. On the other hand, the data-based algorithm, devoid of modeling bias, can improve prediction for both lane-changing maneuvers. We quantify the prediction improvement using ROC curves and demonstrate statistical significance by taking into account the number of parameters.

Automated summary

The study compared the MOBIL model and naive Bayes algorithm for predicting driver lane-changing intents, finding that the data-based algorithm can improve prediction accuracy for lane-changing maneuvers and is not affected by modeling bias.