Vehicle Trajectory Prediction Based on Posterior Distributions Fitting and TCN-Transformer
Heng Yuan, Jun Zhang, Lei Zhang, Zhiqiang Zhang, Zhenpo Wang
Abstract
Accurate and efficient prediction of future motions of surrounding vehicles plays a crucial role in navigating complex traffic scenarios for automated vehicles. To address the poor interpretability of predicted distribution boundaries and the low accuracy of distribution mean with the existing methods, this paper presents a novel deep learning-based approach for predicting future trajectory distributions of surrounding vehicles. The proposed method consists of two key parts: the posterior trajectory distribution module (PTDM) and the Temporal Convolutional Network (TCN)-Transformer module. The PTDM module fits the posterior distribution for each recorded trajectory in the training set, while the TCN-Transformer module models the future trajectory distribution. The model training incorporates the boundaries of the posterior distribution derived from PTDM. To evaluate the proposed scheme, the NGSIM and inD datasets are utilized, and the results show an accuracy improvement of 8.4% compared to the state-of-the-art methods at a prediction horizon of 4 seconds. Furthermore, the proposed method exhibits high computational efficiency, with an average computational time of 0.05 ms and a model size of 1/6th of the GRU-based model.