A local posture manifold with Lagrangian parallel constraint for subtle action discrimination

Abstract

To effectively discriminate the so called “significant motion”, actions with subtle differences, such as minimum inertia among running, jogging and walking, are approximated by a new local posture descriptor. First, each human pose from action sequences is divided into multiple local rigid body-parts (LRBPs) by the multi-group 2-simplex templates. Second, a new local posture descriptor is proposed by a local principal manifold model based on Lagrangian parallel constraint (LPC-LPM) to describe each LRBP. Further, each non-rigid human pose is expressed by summing up all linear weighted probabilities of the geometry distribution of local posture manifolds. In addition, to improve the discrimination performance, spatio-temporal context descriptors of LRBPs are extracted as enhanced features. Experimental results show that our proposed approach achieve higher recognition rate for significant actions, which is better than previously results.