Vehicle Lane-changing Trajectory Prediction Based on Interactive Multiple Model

This study aims to improve the accuracy of trajectory prediction in lane-changing scenarios compared to the state-of-the-art. Lane-changing trajectory prediction is critical for autonomous vehicle driving safety in the complex traffic environment. This paper proposed a novel vehicle lane-changing prediction method by combining kinematics and data-driven-based methods in an interactive framework. Firstly, a kinematics-based prediction method, Constant Angle Rate and Velocity Model (CTRV), and a data-driven prediction method, Long Short-Term Memory Network (LSTM) are systematically compared. It is demonstrated that CTRV is difficult to capture long-horizon lane-changing dynamics, and generates the results in low accuracy. On the contrary, LSTM performs better in lane-changing long-horizon scenarios, but worse in short-horizon scenarios, due to the difficulty of precise fitting in data-driven architecture. In this case, we construct a novel interactive multi-model (IMM) trajectory prediction method that combines the above two prediction models. This method successfully captures the short-horizon and long-horizon vehicle lane-changing dynamics and improves the prediction accuracy greatly. Public dataset NGSIM I-80 is adopted for simulation and validation. Results show that the IMM-based approach reduced the lane-changing trajectory prediction error by 3% at least compared to LSTM methods.