Novel trajectory prediction algorithm using a full dataset: comparison and ablation studies

Recent research efforts into pedestrian-CAV interactions have focused on large-scale, real-world data, eschewing the experimental approaches that are required for controlled testing and investigations. This study introduces a dataset investigating pedestrian-CAV interactions, generated by extensive, networked virtual reality experiments. It then goes on to develop a model that investigates behaviour in road-crossing scenarios and utilises this dataset to train and test the model. We investigate the performance of the model relative to other, state-of-the-art approaches, as well as carrying out an ablative study to investigate the relative importance of the various features obtained within the VR environment. This study finds that our model outperforms previous state-of-the-art approaches, that the most powerful predictive variable when investigating road-crossing is the time-to-collision feature, and that hierarchical fusion techniques should be used when combining contextual scenario features.