TrajNet++: Trajectory forecasting benchmark & challenge

Open data and method: A trajectory forecasting benchmark

Prof. Alexandre Alahi

Reproducible results are a must for any scientific publication.

Prof. Alexandre Alahi, VITA


A machine learning challenge, with accompanying data and evaluation codes, open to the public on All scripts are shared under MIT license.


Why Open?

Most of the recent methods in trajectory forecasting research are not objectively comparable, as they have often been evaluated on different subsets of available data, without proper sampling of trajectories. To allow for an objective comparison of forecasting techniques, Alahi’s team set up a trajectory forecasting challenge TrajNet++ that is open to the public.


Who benefits?

TrajNet++ benchmark benefits researchers working on human motion forecasting, an active area of research for the transportation community. In its first year, the challenge already received almost 1000 submissions from over 200 participants.


The team developed a framework for the fair evaluation of trajectory forecasting algorithms, and provided both a curated dataset, with clear categorisation of trajectories to ensure reproducible sampling, and an extensive evaluation tool, facilitating a fair comparison of performance.