Safety-Aware Vehicle Trajectory Prediction With Spatiotemporal Attentional GAN in Hybrid Transportation System

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; pp. 1 - 13
• Main Authors: Huo, Jie, Wang, Luhan, Wen, Xiangming, Liu, Luning, Yao, Guanyu, Lloret-Batlle, Roger, Lu, Zhaoming
• Format: Journal Article
• Language: English
• Published: New York IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Behavioral sciences; Clustering; Datasets; Driving; Driving pattern clustering; Generative adversarial networks; generative adversarial networks (GANs); Hybrid systems; hybrid transportation system; Prediction models; Predictive models; Root-mean-square errors; Safety; Servers; spatiotemporal attention; Trajectory; Transportation; Transportation systems; vehicle trajectory prediction
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2023.3327472

During the transition from human-driven vehicles (HDVs) to a fully connected automated vehicle (CAV) traffic environment, the hybrid transportation system in which HDVs and CAVs coexist will face challenges owing to the uncertainty of HDVs trajectories. Accurately predicting the trajectories of HDVs and conducting safety warnings for them are essential to improve the safety of the hybrid transportation system. In this article, we propose a safety-aware vehicle trajectory prediction with spatiotemporal attentional generative adversarial network (GAN) (SSTAttGAN) in the hybrid transportation system. Considering multiple driving characteristics, a driving pattern clustering mechanism based on spectral clustering is introduced to analyze vehicle behavior from the perspective of enhancing safety. In the GAN-based trajectory prediction model, a spatiotemporal attention mechanism is proposed as a module of the GAN generator, through which optimal weight distributions of various spatiotemporal features affecting trajectories can be obtained to improve prediction accuracy. The experimental evaluation on the real trajectory dataset demonstrates that our scheme outperforms state-of-the-art methods. Specifically, the root-mean-square error (RMSE) of our model on the next-generation simulation (NGSIM) dataset is 0.39 and 2.09 m in the predicted 1- and 5-s time horizons, respectively. Also, the RMSE on the HighD dataset is 0.09 and 0.92 m in the predicted 1- and 5-s time horizons, respectively.