Study on Small-scale Dataset-based Vehicle Trajectory Prediction in Urban Abnormal Situations

For the autonomous driving safety, it is important to know the future trajectories of surrounding vehicles, especially during the transition period where the autonomous and non-autonomous vehicles are mixed each other. In this regard, neural-network models including the long short-term memory(LSTM) have demonstrated outstanding performance in the field of vehicle trajectory prediction, but their accuracy significantly decreases in the prediction for abnormal situations, because the dataset for them are usually small-scaled and the prediction results tend to be biased to the maneuvers of large portion. To tackle this problem, we propose a trajectory prediction framework that incorporates classification-based switching mechanism. After the maneuver was classified by the radial basis function(RBF) kernel-based support vector machine(SVM), the prediction results were selectively obtained from multiple LSTM-based prediction models in the proposed framework, where each prediction model was trained with dataset for each maneuver. In this way, the future trajectories could be predicted successfully even for the abnormal maneuvers because the small-scale dataset for them could train the model independently in the proposed framework. The proposed framework was trained and validated with the real trajectory dataset collected at Bang-I Station intersection. The sequence of vehicles’ speed, yaw-rate, latitude and longitude coordinate were used as inputs in the proposed framework.