Improving robustness and precision in GEI + HOG action recognition

Abstract

Histograms of Oriented Gradients is a well known and applied descriptor, however "black box" use is common. Gradient computation is the key to performance and may be application dependent. In this paper we examine explicit, implicit and Hessian schemes as opposed to the recommended centred mask. Results indicate the explicit Bickley scheme boosts robustness, both static and dynamic information are important to recognition and full body Gait-Energy Images are preferred. Robustness is boosted by specific choice of cell and bin parameters and SVM where actions are pre-classified using temporal information. © 2013 Springer-Verlag.