Research map / Video / Video segmentation and tracking
Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TRACK, a framework that jointly pre-trains dynamic road segment and trajectory representations by linking traffic state data and trajectory data via transition-aware GAT, traffic transformer, co-attentional encoder and trajectory-traffic state matching.
Problem
Existing road network and trajectory representation learning methods are static and model traffic state and trajectory data independently, missing their mutual dynamic influences needed for downstream traffic tasks.
Approach
Incorporates trajectory-derived time-aware transition probabilities into GAT attention, models traffic states with a spatio-temporal traffic transformer, and exchanges information across the two views via a co-attentional transformer with gravitation-inspired distance-decayed attention, aligned by a trajectory-traffic state matching contrastive task.
Key results
- On multi-step traffic state prediction, TRACK reaches MAE 1.094 and MAPE 13.32% on Xi'an, beating the best baseline SSTBAN (1.175, 14.11%).
- On travel time estimation, TRACK reaches MAE 1.426 on Xi'an versus 1.522 for START and MAE 1.143 vs 1.182 on Chengdu.
- Ablations show removing traffic-state or trajectory modeling each degrades both downstream tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRACK instead of SSTBAN | Multi-step traffic state prediction · Xi'an | MAE: 1.175 -> 1.094 | Helped |
| Use TRACK instead of MTGNN | Multi-step traffic state prediction · Chengdu | MAPE (%): 16.76 -> 15.42 | Helped |
| Use TRACK instead of START | Travel time estimation · Xi'an | MAE: 1.522 -> 1.426 | Helped |
| Use TRACK instead of START | Travel time estimation · Chengdu | RMSE: 1.749 -> 1.612 | Helped |
| Add traffic-state dynamics modeling (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
| Add trajectory representation learning and trajectory-based dynamic spatial features (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
| Add trajectory-traffic state matching (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
| Add timestamp masking prediction loss (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
| Add segment masking prediction loss (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
| Add contrastive trajectory learning (ablation) | TRACK · Xi'an MSTSP and TTE | performance | Helped |
Limitations
- Evaluated on only two cities and two downstream tasks.
- Relies on the DiDi GAIA dataset, limiting generality of conclusions.
- Hardware and compute cost are not reported.
Experimental setup
Datasets: DiDi GAIA Xi'an trajectories and traffic states DiDi GAIA Chengdu trajectories and traffic states
Benchmarks: Xi'an DiDi GAIA Chengdu DiDi GAIA
Metrics: MAE MAPE RMSE
Baselines: DCRNN GWNET MTGNN TrGNN STGODE ST-Norm SSTBAN traj2vec t2vec Trembr PIM Toast
Builds on: graph attention networks transformer NT-Xent contrastive loss START
Closest papers by meaning
- Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting (AAAI 2019 · similarity 0.89)
- Traffic Flow Prediction with Vehicle Trajectories (AAAI 2021 · similarity 0.88)
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.87)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.87)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.86)
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.