GEI + HOG for Action Recognition

This paper demonstrates the benefit of applying the Gait-Energy Image (GEI) [15] and Histograms of Oriented Gradients (HOG) [12] descriptors for action recognition. Multi-class Support Vector Machine (SVM) classification show promising results at 100% using leave-one-out cross validation. Furthermore, this technique gains 27◦ viewpoint tolerance and robustness to occlusions, clothing and carrying condition variations. The contribution of this paper is two-fold. The first employs a traditional gait recognition representation alongside HOG descriptors for action recognition, while the second decomposes actions into static and dynamic classes for superior performance and reduced processing time.