K-Subspaces for Sequential Data
Wubin Sheng, John J. Lipor · 2023
We study the problem of clustering high-dimensional temporal data such as video sequences of human motion, where points that arrive sequentially in time are likely to belong to the same cluster. State-of-the-art approaches to this problem rely on the union-of-subspaces model, where points lie near one of$K$unknown low-dimensional subspaces. We propose the first approach to sequential subspace clustering based on the popular$K$-Subspaces (KSS) formulation, which we refer to as Temporal$K$-Subspaces (TKSS). We show how sequential information can be incorporated into the KSS problem and provide an efficient algorithm for approximate minimization of the resulting cost function, proving convergence to a local minimum. Results on benchmark datasets show that TKSS achieves state-of-the-art performance, obtaining an accuracy increase of over 10% compared to existing methods.