Data-Centric Model for Multiclass Lane Change Maneuver Prediction Using a Naturalistic Driving Dataset

Connected and automated vehicles (CAVs) are rapidly becoming a reality. It is easy to hypothesize they will someday become the dominant vehicle on the highway system. However, before this occurs, CAVs and human-driven vehicles (HDVs) will experience numerous conditions that generate traffic conflicts in the mixed flow context during the transition time. Therefore, it is important to identify and predict the driving intentions of HDV in order to mitigate or eliminate traffic conflicts and vehicle collisions. The current highway maneuver prediction models do not do well at multiclass lane change maneuver (LCM) detection and prediction near ramp areas. In this paper, an LSTM neural network for predicting LCM, using data from the ExiD datasets is developed and proposed. The ExiD dataset is a trajectory-based dataset collected on German highway ramps using unmanned aerial vehicles. The proposed model will use a long short-term memory (LSTM) neural network to detect the LCM and trained for different values of prediction horizon time ranging from 0.5 s to five seconds. As will be shown the proposed approach, the LCM detection problem is converted into an anomaly “sawtooth” pattern detection. The results of the proposed model show the accuracy rate of the proposed LCM algorithms is 98% when predicting 0.5 s ahead. The false alarm rate is relatively small at 5.4%. The LCM detector, which uses the “sawtooth” pattern of the lateral shift distance variable, can be used to predict double LCMs. It is hypothesized this would benefit in detection and prediction of overtaking and risky cut-in and cut-out driving behaviors. Ultimately this would enhance highway safety in the vicinity of the on and off ramp sections.