Research map / Probabilistic and scientific modeling / Molecules and proteins
Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking
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.
Introduces TrajImpute, a pedestrian trajectory dataset with simulated missing observed coordinates (easy/hard protocols on ETH/HOTEL/UNIV/ZARA1/ZARA2), benchmarks six imputation methods and six trajectory predictors on imputed data.
Problem
Pedestrian trajectory prediction methods and datasets assume fully observed past trajectories, ignoring missing coordinates caused by sensor failure, occlusion, and limited field of view.
Approach
A dataset that injects missing coordinates into observed trajectories of standard pedestrian benchmarks under easy (0–4 missing of 8) and hard (4–7 missing) protocols, with uniform random missingness, paired with a unified benchmark for imputation followed by trajectory prediction.
Key results
- SAITS achieves the lowest imputation MAE on most subsets, e.g., ETH-M easy MAE 0.5031 vs. 0.6467 for US-GAN (next best).
- Trajectory prediction degrades under hard imputation for all models; e.g., GraphTern on ETH goes from clean 0.42/0.58 to 0.78/0.77 ADE/FDE with hard-impute.
- EQMotion shows the smallest degradation under hard imputation (ETH 0.47/0.63 vs. clean 0.40/0.61).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · ETH-M easy impute | MAE: 0.6467 -> 0.5031 | Helped |
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · HOTEL-M easy impute | MAE: 2.6327 -> 2.1930 | Helped |
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · UNIV-M easy impute | MAE: 0.6713 -> 0.1939 | Helped |
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · ZARA1-M easy impute | MAE: 0.3125 -> 0.2054 | Helped |
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · ZARA2-M easy impute | MAE: 0.2409 -> 0.1314 | Helped |
| Use SAITS instead of the strongest competing imputation method | Trajectory imputation · ZARA1-M hard impute | MAE: 0.5699 -> 0.6277 | Hurt |
| Increase missing observations from easy to hard protocol | Imputation methods · TrajImpute | MAE | Hurt |
| Use hard-imputed rather than clean observations | Trajectory prediction models · ETH, Hotel, UNV, ZARA1, and ZARA2 | ADE/FDE | Hurt |
| Use easy-imputed rather than clean observations | GraphTern trajectory prediction · ETH | ADE/FDE: 0.42/0.58 -> 0.77/0.74 | Hurt |
| Use easy-imputed rather than clean observations | GraphTern trajectory prediction · Hotel | ADE/FDE: 0.14/0.23 -> 0.15/0.25 | Mixed |
Surprising findings
- On ZARA1 hard protocol, TimesNet slightly outperforms SAITS in imputation MAE (0.5699 vs. 0.6277).
- Some trajectory predictors perform better on easy-imputed than on clean data for certain subsets (e.g., LBEBM-ET on ETH: 0.37/0.55 easy-impute vs. 0.36/0.53 clean is similar; SGCN-ET and others degrade).
Limitations
- Only five standard ETH/UCY-style datasets with world-2D coordinates are used; other datasets are not simulated.
- Existing imputation and prediction methods are not designed for missing observed coordinates, and performance drops sharply under hard protocol.
- Missingness is synthetic and uniform random rather than from real sensor failures or occlusion.
Experimental setup
Datasets: ETH HOTEL UNIV ZARA1 ZARA2
Benchmarks: TrajImpute ETH-M TrajImpute HOTEL-M TrajImpute UNIV-M TrajImpute ZARA1-M TrajImpute ZARA2-M
Metrics: MAE MSE RMSE MRE ADE FDE
Baselines: Transformer US-GAN BRITS M-RNN TimesNet GraphTern LBEBM-ET SGCN-ET EQMotion TUTR GPGraph
Builds on: SAITS BRITS M-RNN TimesNet
Closest papers by meaning
- MS-TIP: Imputation Aware Pedestrian Trajectory Prediction (ICML 2024 · similarity 0.92)
- ProDiff: Prototype-Guided Diffusion for Minimal Information Trajectory Imputation (ICML 2025 · similarity 0.88)
- TrajCLIP: Pedestrian trajectory prediction method using contrastive learning and idempotent networks (NeurIPS 2024 · similarity 0.88)
- Towards Predicting Any Human Trajectory In Context (NeurIPS 2025 · similarity 0.88)
- SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents (AAAI 2024 · similarity 0.87)
- Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion (ICML 2025 · similarity 0.87)
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.