New Method Quantifies Social Interaction for Pedestrian Path Prediction.

Accurate long-term human path forecasting in crowded environments is vital for the safe and efficient operation of autonomous platforms, such as self-driving cars and social robots. While existing research considers social interactions in predictions, it often fails to explicitly identify the types of interactions occurring or how they influence pedestrian decision-making, which limits the robustness of these models. This research creatively addresses this gap by proposing "Learn to Cluster," a novel method to quantify and interpret pedestrian social interactions. This approach is a label-free, probabilistic latent variable generative model that learns directly from sequential trajectory observations, making it scalable to an arbitrary number of pedestrians. The latent variables discovered by the clustering process then serve as 'labels' to categorize distinct social interactions. Crucially, "Learn to Cluster" can be seamlessly integrated into the training process of trajectory prediction models. Extensive experiments conducted on several trajectory prediction benchmarks demonstrate that the method successfully learns social interaction patterns and effectively incorporates them into pedestrian trajectory prediction, leading to improved forecasting accuracy.