Invariants for motion-based classification

Abstract

Shape geometric invariants play an important role in model-based vision (MBV). However, in many MBV scenarios, shape information may not be sufficiently reliable (e.g., camouflage or concealment) and hence other types of invariants need to be considered. The paper addresses motion-based classification of objects based on unique motion or activity characteristics in a long-sequence of images. To date, the techniques developed in motion-based recognition are inherently sensitive to (a) object's shape, (b) Euclidean group actions and (c) time scale, i.e., velocity and acceleration of motion. We propose the development of a set of motion-based invariants that capture geometric aspects of an object's kinematic constraints during distinctive motions and activities. Algebraic and differential invariants of curves and surfaces in a projective space, the kinematic image space, are proposed for motion and activity classification. The proposed approach establishes parallelism between shape and motion geometric invariance.