Motion Prediction Challenge

The future brings forth a promise to make transportation safer and easier with the advent of Autonomous systems. In order to successfully navigate through diverse scenarios, these systems need to understand the rules of human motion regarding social interactions and physical interactions. These social etiquettes come intuitively to humans through years of observation. A task closely related to understanding human motion is to forecast the movement of the surrounding people, conforming to the common sense unspoken rules of human motion as well as the surrounding physical constraints. We refer to this task as trajectory forecasting.

In this workshop, we will run a challenge on trajectory forecasting focussed on modeling agent-agent interactions referred to as our TrajNet++ challenge. The challenge will be running for the first time in the upcoming AMLD'20 conference in January 2020. It will be in its second edition in ICRA 2020.

TrajNet++ Challenge

• Agent-Agent: In the past few years, several novel methods have been proposed to tackle agent-agent interactions. However, most methods have been evaluated on limited data. Furthermore, these methods have been evaluated on different subsets of the available data without proper indexing of trajectories making it difficult to objectively compare the forecasting techniques. In order to tackle this issue, we propose a large-scale framework that provides not only proper indexing of trajectories but also a unified extensive evaluation system to test the gathered methods for a fair comparison. In this agent-agent challenge, researchers have to study how their method performs in explicit agent-agent scenarios.
• Agent-Space: To anticipate motion of each agent, scene should also be taken into account due to its strong constraint. Despite its importance, few studies in the literature targeted scene and most of them focused on agent-agent interaction. Moreover, accurate annotated dataset rich in agent-scene interaction is rare. We propose another sub-challenge to address agent-scene interaction. Researchers have to study how scene constraint can be added to the models in a generalizable way. We provide a new EPFL-roundabout dataset which consists of 4 roundabouts rich in vehicle-scene interactions. We hope this challenge leads to progress in the agent-scene interaction problem which will advance current research on trajectory prediction.

How to partecipate (on-going)

• Trajnet++ (A Trajectory Forecasting Challenge) website
• More details can be found on here

Organizers

• Parth Kothari, EPFL
• Sven Kreiss, EPFL
• Alexandre Alahi, EPFL