Human Trajectory Forecasting in Crowds: A Deep Learning Perspective

Trajectory forecasting in crowded scenes has become an important topic in recent times because of its increasing demand in emerging applications like autonomous cars, service robots, intelligent tracking. One crucial challenge in trajectory forecasting is to effectively model the social interactions between different agents. In the past few years, several novel methods have been proposed to model agent-agent interactions. However, (1) these methods have been evaluated on different subsets of the available data without a proper sampling of interaction-centric trajectories (2) the current evaluation system lacks metrics that capture the social feasibility of model predictions. This makes it difficult to objectively compare the various forecasting methods.

We introduce TrajNet++, a large-scale interaction-centric trajectory forecasting benchmark where researchers can evaluate how their method performs in explicit agent-agent scenarios. Our benchmark provides not only a proper sampling of socially interacting trajectories but also an extensive evaluation system with novel metrics to test the gathered methods for a fair comparison.