Self-Supervised Trajectory Clustering with Optimal Transport and Metric Learning

Unraveling the complex patterns embedded in tra-jectory data offers profound insights into various applications, from urban planning to environmental impact assessments. This paper introduces a self-supervised trajectory clustering framework that synthesizes the strengths of deep learning models with advanced loss functions to address the nuances of spatial-temporal relationships. Our model integrates the Sinkhorn Distance to leverage the geometric nature of trajectories, enabling the optimal transport of mass in embedding space. Further, we employ a modified Generalized End-to-End (GE2E) loss, typically used in speaker verification, to fine-tune the latent space for distinct clustering results without reliance on explicit labels. Our results significantly outperform existing deep clustering methods, underscoring the potential of our approach in capturing the intricate characteristics of trajectory data for high-quality clustering outcomes and practical environmental applications, by identifyina movement patterns.