Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UniTraj: A Unified Framework for Scalable Vehicle Trajectory Prediction

Lan Feng, Mohammadhossein Bahari, Kaouther Messaoud Ben Amor, Éloi Zablocki, Matthieu Cord, Alexandre Alahi

TL;DR

UniTraj addresses the challenge of cross-domain generalization in vehicle trajectory forecasting by unifying datasets, models, and evaluation. The framework standardizes data formats via ScenarioNet, harmonizes features across diverse datasets, and provides a common evaluation suite, enabling rigorous cross-dataset and cross-city experiments. Empirical results show significant generalization gaps across datasets, but data scaling and increased diversity yield substantial performance gains, achieving state-of-the-art results on nuScenes when trained on all included data. By delivering an open-source platform with comprehensive dataset analyses and support for multiple models, UniTraj facilitates robust, cross-domain trajectory prediction research and practical deployment in heterogeneous driving environments.

Abstract

Vehicle trajectory prediction has increasingly relied on data-driven solutions, but their ability to scale to different data domains and the impact of larger dataset sizes on their generalization remain under-explored. While these questions can be studied by employing multiple datasets, it is challenging due to several discrepancies, e.g., in data formats, map resolution, and semantic annotation types. To address these challenges, we introduce UniTraj, a comprehensive framework that unifies various datasets, models, and evaluation criteria, presenting new opportunities for the vehicle trajectory prediction field. In particular, using UniTraj, we conduct extensive experiments and find that model performance significantly drops when transferred to other datasets. However, enlarging data size and diversity can substantially improve performance, leading to a new state-of-the-art result for the nuScenes dataset. We provide insights into dataset characteristics to explain these findings. The code can be found here: https://github.com/vita-epfl/UniTraj

UniTraj: A Unified Framework for Scalable Vehicle Trajectory Prediction

TL;DR

UniTraj addresses the challenge of cross-domain generalization in vehicle trajectory forecasting by unifying datasets, models, and evaluation. The framework standardizes data formats via ScenarioNet, harmonizes features across diverse datasets, and provides a common evaluation suite, enabling rigorous cross-dataset and cross-city experiments. Empirical results show significant generalization gaps across datasets, but data scaling and increased diversity yield substantial performance gains, achieving state-of-the-art results on nuScenes when trained on all included data. By delivering an open-source platform with comprehensive dataset analyses and support for multiple models, UniTraj facilitates robust, cross-domain trajectory prediction research and practical deployment in heterogeneous driving environments.

Abstract

Vehicle trajectory prediction has increasingly relied on data-driven solutions, but their ability to scale to different data domains and the impact of larger dataset sizes on their generalization remain under-explored. While these questions can be studied by employing multiple datasets, it is challenging due to several discrepancies, e.g., in data formats, map resolution, and semantic annotation types. To address these challenges, we introduce UniTraj, a comprehensive framework that unifies various datasets, models, and evaluation criteria, presenting new opportunities for the vehicle trajectory prediction field. In particular, using UniTraj, we conduct extensive experiments and find that model performance significantly drops when transferred to other datasets. However, enlarging data size and diversity can substantially improve performance, leading to a new state-of-the-art result for the nuScenes dataset. We provide insights into dataset characteristics to explain these findings. The code can be found here: https://github.com/vita-epfl/UniTraj
Paper Structure (29 sections, 1 equation, 4 figures, 17 tables)

This paper contains 29 sections, 1 equation, 4 figures, 17 tables.

Table of Contents

  1. Introduction
  2. Previous work
  3. UniTraj framework
  4. Unified data
  5. Unified data format:
  6. Unified data features:
  7. Unified models
  8. Unified evaluation
  9. General evaluations:
  10. Experiments
  11. Generalization evaluation
  12. Cross-dataset evaluation:
  13. Cross-city evaluation:
  14. Scaling data to 2M trajectories.
  15. Analyzing the results
  16. ...and 14 more sections

Figures (4)

  • Figure 1: UniTraj framework. The framework unifies various datasets, forming the largest vehicle trajectory prediction dataset available. It also includes multiple state-of-the-art prediction models and various evaluation strategies, making it suitable for trajectory prediction experimentation. The framework enables the study of diverse Research Questions, including (RQ1) the generalization of trajectory prediction models across different domains and (RQ2) the impact of data size on prediction performance.
  • Figure 2: Relationship between dataset size and model performance. The prediction error of AutoBot as the combined dataset size increases, varying from 20% to 100% of the total data.
  • Figure 3: Figure (a) shows the distribution of trajectory types. It reveals an imbalance across different types with straight being the most common trajectory type in the datasets. Figure (b) shows the histogram of the Kalman Difficulty of trajectories. To give a sense of the Kalman difficulty, we overlay three random examples. The past trajectory, the ground truth, and the Kalman filter prediction are shown in red, blue, and magenta, respectively. The plot displays a clear trend with a notably higher count of simpler scenarios compared to challenging ones. WOMD, in particular, shows a relatively balanced distribution across scenarios.
  • Figure 4: nuScenes Leaderboard. We train AutoBot and MTR with all datasets, and evaluate on nuScenes (ranking at the time of submission among public methods)