Vehicle Trajectory Prediction for Automated Driving Based on Temporal Convolution Networks

Abstract

For automated driving, trajectory prediction of other surrounding vehicles is crucial to ego vehicle’s driving decision. This is especially important when automated vehicles, e.g. SAE level 3 vehicles or fully-automated robotaxis, share the open road with human-driven vehicles. One of the main research directions in this field is to adopt the methods of deep learning. We propose a TCN-MLP encoder-decoder framework that considers both the trajectory data of predicted vehicle and surrounding vehicles in the input. To handle more complex trajectory prediction, a driving intention recognition module is added to the model to identify the intentions in longitudinal and lateral directions. Based on the HighD dataset, we have tested the proposed model and its two variations for ablation experiment. The results show that our model can achieve more accurate trajectory prediction than the state-of-the-art approaches, and the prediction RMSE is reduced by about 33.3% on average. Our model may serve as a promising solution to vehicle trajectory prediction problems in highway scenes.