Trustworthy Pedestrian Trajectory Prediction via Pattern-Aware Interaction Modeling

The introduction of InSyn, a Transformer-based model for pedestrian trajectory prediction, marks a significant advancement in the field of artificial intelligence. Traditional methods often relied on black-box approaches that compromised the reliability of predictions due to their opaque nature. InSyn overcomes this limitation by explicitly modeling diverse interaction patterns among pedestrians, such as walking in sync or in conflict. This transparency is crucial for applications in crowded environments where accurate predictions are essential. The model's innovative training strategy, Seq-Start of Seq (SSOS), effectively addresses the common issue of initial-step divergence in time-series predictions, leading to an approximate 6.58% reduction in prediction errors. Experiments conducted on the ETH and UCY datasets demonstrate that InSyn not only surpasses recent black-box baselines in accuracy, particularly in high-density scenarios, but also provides a clearer understanding of pedest…