A Perceptually-Inspired Stochastic Framework for Video Search and Analysis
Ricky J. Sethi · 2014
The neurobiological model for motion recognition and visual processing posits that visual stimuli in the brain bifurcate into a Motion Energy Pathway and a Form Pathway, both of which are finally Integrated in the resulting motion recognition. In this article, we propose a perceptually-inspired computational framework for video search and analysis of human activity recognition. For high-resolution video, our computational framework relies upon the physics-based Human Action Image for the Motion Energy Pathway, shapes of trajectories and Dynamic Time Warping for the Form Pathway, and a variant of the bootstrap for the Integration. For low-resolution video, the computational equivalent uses simple, physics-based kinematic features for the Motion Energy Pathway, Space-time Interest Points and Nearest Neighbour classification for the Form Pathway, and a variant of the Markov chain Monte Carlo method for the Integration. We demonstrate the efficacy of our system on real-life video sequences from the well-known USF, Weizmann, UCR Videoweb, and YouTube/Hollywood Human Actions datasets.