Group action induced distances for averaging and clustering Linear Dynamical Systems with applications to the analysis of dynamic scenes

Abstract

We introduce a framework for defining a distance on the (non-Euclidean) space of Linear Dynamical Systems (LDSs). The proposed distance is induced by the action of the group of orthogonal matrices on the space of statespace realizations of LDSs. This distance can be efficiently computed for large-scale problems, hence it is suitable for applications in the analysis of dynamic visual scenes and other high dimensional time series. Based on this distance we devise a simple LDS averaging algorithm, which can be used for classification and clustering of time-series data.We test the validity as well as the performance of our groupaction based distance on synthetic as well as real data and provide comparison with state-of-the-art methods.

Bio-inspired Dynamic 3D Discriminative Skeletal Features for Human Action Recognition

“…Over the years, several metrics have been proposed e.g. [8,17,25,1,6]. Of these, the Martin distance has been the most extensively used as it is invariant to the noise statistics as well as initial state of the dynamical system.…”

Intrinsic Characterization of Dynamic Surfaces

“…Hence, clustering algorithms used in the Euclidean space (such as k-means) cannot be applied directly, as discussed in prior work [7,5,6,1].…”

Recognizing Activities via Bag of Words for Attribute Dynamics

“…The main challenge of this dictionary leaning problem is the difficulty of identifying the "centroid" of a collection of dynamic textures, due to the non-Euclidean nature of the space of linear dynamic systems. [20] bypasses this problem with resort to a somewhat heuristic combination of multi-dimensional scaling and kmeans (denoted MDS-kM); while [1] presents a procedure to directly average dynamic models in the parameter space, the approach only works for LDS's. We propose an alternative principled solution, which is specifically designed for clustering attribute sequences, and has a number of advantages over MDS-kM.…”

“…This is related to the bag-of-systems framework of [20,1], where a set of dynamic textures (DTs) [5] were used to characterize dynamic scenes. The main challenge of this dictionary leaning problem is the difficulty of identifying the "centroid" of a collection of dynamic textures, due to the non-Euclidean nature of the space of linear dynamic systems.…”