Core Concepts
VT-Former combines transformers and graphs for accurate vehicle trajectory prediction in highway surveillance.
Abstract
1. Abstract:
Vehicle Trajectory Prediction (VTP) is crucial for road safety.
VT-Former introduces a transformer-based approach with Graph Attentive Tokenization (GAT).
2. Introduction:
ITS distinguishes between autonomous driving and surveillance applications.
VTP is essential for traffic management, accident prevention, and more.
3. Challenges in SVTP:
Balancing efficiency and accuracy is crucial.
Predicting vehicle interactions is complex.
4. VT-Former Methodology:
Utilizes transformers and GAT for accurate trajectory prediction.
5. Experimental Setup:
Evaluation on NGSIM, CHD High-angle, and CHD Eye-level datasets.
6. Results:
VT-Former outperforms SotA models in trajectory prediction accuracy.
7. Observation Horizon Analysis:
Shorter observation horizons improve prediction accuracy, especially in eye-level scenarios.
8. Future Directions:
Explore other graph networks and evaluate performance under resource constraints.
Stats
VT-FormerはNGSIMデータセットでADEが0.81メートル、FDEが1.80メートルを達成しました。