Combining Domain Knowledge and Deep Learning Methods for Vehicle Trajectory Prediction

Abstract

Predicting the future trajectory of surrounding agents is especially crucial for autonomous vehicles applied in dense traffic streams. Majority of the approaches presently implemented for vehicle trajectory prediction can be generally classified into domain knowledge-driven method and deep learning approach. Although domain priori knowledge such as traffic rules implementing in knowledge-driven method has realistic output, the interactive performance with other traffic agents is constrained. Conversely, data-driven approach can acquire superior interactive performance by training the model with huge amounts of data, but the trajectory prediction cannot completely satisfy kinematic constraints, especially to datasets the model has not been explicitly trained on. After carefully design and verification, we put forward a method that combining domain knowledge and deep learning model to output accurate and realistic trajectory prediction. Lastly, we evaluate the result on a public available dataset.