Robust Long-Term Vehicle Trajectory Prediction Using Link Projection and a Situation-Aware Transformer

• Published in: Sensors (Basel, Switzerland) Switzerland), 2024-04, Vol.24 (8), p.2398
• Main Authors: Kim, Minsung, Kwak, Byung Il, Hou, Jong-Uk, Kim, Taewoon
• Format: Article
• Language: English
• Subjects: Accuracy; Datasets; Deep learning; Electric transformers; Forecasting; Forecasts and trends; intelligent transport system; Neural networks; Performance evaluation; predictive model; Roads & highways; situation-aware transformer; Time series; Traffic control; trajectory prediction; Vehicles
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s24082398

The trajectory prediction of a vehicle emerges as a pivotal component in Intelligent Transportation Systems. On urban roads where external factors such as intersections and traffic control devices significantly affect driving patterns along with the driver's intrinsic habits, the prediction task becomes much more challenging. Furthermore, long-term forecasting of trajectories accumulates prediction errors, leading to substantially inaccurate predictions that may deviate from the actual road. As a solution to these challenges, we propose a long-term vehicle trajectory prediction method that is robust to error accumulation and prevents off-road predictions. In this study, the Transformer model is utilized to analyze and forecast vehicle trajectories. In addition, we propose an extra encoding network to precisely capture the effect of the external factors on the driving pattern by producing an abstract representation of the situation nearby the driver. To avoid off-road predictions, we propose a post-processing method, called link projection, which projects predictions onto the road geometry. Moreover, to overcome the limitations of Euclidean distance-based evaluation metrics in evaluating the accuracy of the entire trajectory, we propose a new metric called area-between-curves (ABC). It measures the similarity between two trajectories, and thus the accordance between the two can be effectively evaluated. Extensive evaluations are conducted using real-world datasets against widely-used methods to demonstrate the effectiveness of the proposed approach. The results show that the proposed approach outperforms the conventional deep learning models by up to 65.74% (RMSE), 60.13% (MAE) and 91.45% (ABC).