Overview

As road traffic environments become increasingly complex, protectingvulnerable road users (VRUs), who are susceptible to injury or face elevated risk,has become an important topic in traffic-safety research.VRUs generally refer to road users who are more vulnerable in traffic interactions, including non-motorized road users and pedestrians. A deep understanding and accurate prediction of their behavioral patterns are essential to improving traffic safety. In complex road environments, VRUs face many potential risks, such as traffic-rule violations and high-speed movement, which can lead to serious injuries or collisions. Accurately identifying and predicting VRU trajectories, followed by rigorous evaluation, is therefore a core task for road safety.

In recent years, trajectory prediction has become a key research direction for protecting VRUs. Yet the high randomness of their movement patterns, complex interactions, and abrupt changes in motion state create significant challenges. Although many prediction models have been proposed, fragmented datasets, inconsistent evaluation standards, and incomplete scenario coverage still limit technical progress and cross-domain comparison.

To address these issues, this leaderboard provides comprehensive scenario coverage and unified evaluation standards to assess how effectively trajectory prediction methods protect VRUs and reduce the potential risks autonomous-driving technology poses to vulnerable groups.

Leaderboard

How to Participate

Dataset Introduction

Submit Results

Contestants must submit a prediction-result file and a confidence file, both in .npy format, named test_data_y_{name}_{number_of_modes}.npy and conf_{name}_{number_of_modes}.npy, respectively.

The prediction data are in a five-dimensional matrix with the following dimensions:scenario ID,number of modes,individual ID, time step, feature value. Each feature value includes:

• ‘world_x’: vehicle x-axis position in the global coordinate system
• ‘world_y’: vehicle y-axis position in the global coordinate system

This data structure ensures that the predicted positions of all vehicles at different time steps in each scenario can be accurately recorded, while supporting evaluation of multimodal prediction. To ensure rigor, the number of modes is set to a maximum of 10.

The confidence data are in a three-dimensional matrix with the following dimensions:scenario ID,number of modes,individual ID, used to represent the predicted probability of a trajectory. The confidence of each mode is between 0 and 1; there is no requirement for confidence values to sum across modes, and they may be different or identical.

Results are collected through a form. Participants can clickhere to access the form, complete the form to submit results, and receive analysis and statistics within no more than seven working days, after which the leaderboard will be updated.

Evaluation metrics

Conclusion and Outlook

This trajectory prediction leaderboard aims to provide a unified evaluation platform for researchers of trajectory prediction models in motor vehicle–non-motor vehicle interaction scenarios. Its release enables a deeper understanding of how current models perform in the same task scenarios and promotes continued development of related research. The current version still has limitations, such as insufficient scenario diversity, a lack of label information, an incomplete evaluation system, and a single testing approach. Future versions will continually improve the datasets, evaluation system, and test methods to provide researchers with more comprehensive and accurate evaluation tools and further promote innovation and development in autonomous-driving VRU behavior modeling and prediction.