Abstract
Long-term human path forecasting in crowds is critical for autonomous moving platforms (like autonomous driving cars and social robots) to avoid collision and make high-quality planning. Although current research considers social interactions for prediction, they do not reveal the exact types of social interactions among people and how these interactions affect pedestrians' decision-making processes, limiting robustness.
Social interactions in pedestrian walking are intuitively massive and hard to label and quantify. In this paper, we creatively explore how to quantify and interpret pedestrian interactions by proposing Learn to Cluster. Our clustering of social interactions is probabilistic latent variable generative, learning directly from sequential trajectory observations and scalable to an arbitrary number of pedestrians.
Learn to Cluster is label-free and can be naturally integrated into the training process of the prediction model. The latent variables will then serve as 'labels' to categorize social interactions. Extensive experiments across several trajectory prediction benchmarks demonstrate that our method can learn the patterns of social interactions and effectively integrate them into pedestrian trajectory prediction.
Blogger's Review: This research significantly enhances pedestrian trajectory prediction accuracy through a label-free learning approach, highlighting the importance of social interactions in path forecasting. Its innovation lies in quantifying complex social behaviors into a trainable model, making it a noteworthy advancement for future autonomous driving systems.