MSTT: A Multi-Spatio-Temporal Graph Attention Model for Pedestrian Trajectory Prediction
Accurate prediction of pedestrian movements is vital for autonomous driving, smart transportation, and human–computer interactions. To effectively anticipate pedestrian behavior, it is crucial to consider the potential spatio-temporal interactions among individuals. Traditional modeling approaches o...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-08-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/25/15/4850 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurate prediction of pedestrian movements is vital for autonomous driving, smart transportation, and human–computer interactions. To effectively anticipate pedestrian behavior, it is crucial to consider the potential spatio-temporal interactions among individuals. Traditional modeling approaches often depend on absolute position encoding to discern the positional relationships between pedestrians. Unfortunately, this method overlooks relative spatio-temporal relationships and fails to simulate ongoing interactions adequately. To overcome this challenge, we present a relative spatio-temporal encoding (RSTE) strategy that proficiently captures and analyzes this essential information. Furthermore, we design a multi-spatio-temporal graph (MSTG) modeling technique aimed at modeling and characterizing spatio-temporal interaction data across several individuals over time and space, with the goal of representing the movement patterns of pedestrians accurately. Additionally, an attention-based MSTT model has been developed, which utilizes an end-to-end approach for learning the structure of the MSTG. The findings indicate that an understanding of an individual’s preceding trajectory is crucial for forecasting the subsequent movements of other individuals. Evaluations using two challenging datasets reveal that the MSTT model markedly outperforms traditional trajectory-based modeling methods in predictive performance. |
|---|---|
| ISSN: | 1424-8220 |