Vehicle Trajectory Prediction by Integrating Data-Driven and Knowledge-Guided Technique

In the context of autonomous driving, acquiring the trajectories of surrounding vehicles in advance by autonomous vehicles is a crucial factor in ensuring high-level road safety. While trajectory prediction methods based on deep learning have achieved promising results, these data-driven models lack interpretability and transparency, making their reliable use a significant challenge. In this paper, firstly, an intention-aware spatial-temporal attention network-based trajectory prediction model is constructed, which considers the coupling of driving intention and the interaction with surrounding vehicles, extracts important feature information of vehicles in both temporal and spatial dimensions. Secondly, a vehicle trajectory prediction method via the integration of data-driven and knowledge-guided is proposed, considering both hard and soft constraints. A hard constraint of vehicle kinematics is incorporated into the intention-aware spatial-temporal attention network prediction model to generate physically feasible predicted trajectories and to make this part of the network structure have a human-understandable physical meaning. In addition, by leveraging knowledge related to traffic rules, an auxiliary loss function based on knowledge constraint penalties is designed as a soft constraint to optimize the training of the model and improve the interpretability of the training process. Finally, the proposed model is experimentally evaluated on the datasets and the prediction results are analyzed in terms of reliability and accuracy. The experimental results demonstrate that knowledge guidance effectively enhance the reliability and interpretability of the prediction, and improve the accuracy of long-term trajectory prediction.