Transformer-Based Vehicle-Trajectory Prediction at Urban Low-Speed T-Intersection

Area of Science:

• Artificial Intelligence
• Computer Vision
• Transportation Engineering

Background:

• Transformer models excel at trajectory prediction but require significant computational resources and struggle with long-term predictions.
• Vehicle trajectory prediction is crucial for intelligent transportation systems and autonomous driving, especially in complex urban environments.

Purpose of the Study:

• To develop and optimize a lightweight transformer model for accurate vehicle trajectory prediction in low-speed urban T-intersections.
• To identify optimal model parameters, including loss function and sequence length, for enhanced prediction accuracy and efficiency.
• To evaluate the model's generalization capabilities in atypical driving scenarios and assess the impact of additional features.

Main Methods:

• Generated microscopic traffic simulation data for training and validation, including atypical scenarios.
• Explored various loss functions, settling on smooth L1 loss for optimal performance.
• Examined input/output sequence lengths, determining 1s input and 3s output as optimal.
• Evaluated model generalization using diverse driving-characteristic data.

Main Results:

• The smooth L1 loss function significantly improved prediction accuracy.
• An optimal input sequence length of 1 second and an output sequence length of 3 seconds were identified.
• Enhancing model structure proved more effective than diversifying training data for generalization in atypical situations.
• Incorporating additional features like speed variation reduced model accuracy by approximately 21%.

Conclusions:

• Optimized transformer models offer a viable solution for lightweight trajectory prediction in edge computing environments.
• The findings support the development of trajectory prediction and accident analysis systems for various urban driving scenarios.
• Focusing on model architecture improvements is key to achieving robust performance in complex and atypical driving conditions.