Accurately predicting human motion is a challenging task for assisted or automated vehicles, as humans have latent goals, independent movement styles, are stochastic, and influenced by multiple factors. In recent years, machine learning based approaches have outperformed classical approaches by a large extent. However, existing models are mostly evaluated with general trajectory prediction metrics, neglecting the fact that many metrics are not sufficiently measuring effects of multi-modality, critical situations, and downstream functions.

The 2nd Workshop on Benchmarking Trajectory Forecasting Models at ICCV 2021 was bringing together researchers to tackle problems of proper benchmarks for human motion prediction. We were joining the workshop and presenting results of our latest publication on “Pedestrian Behavior Prediction for Automated Driving: Requirements, Metrics, and Relevant Features”, with a focus on a new system-driven metric for assessing performance of pedestrian prediction for comfortable urban driving.