Visual words selection for human action recognition using rbfnn via the minimization of L-GEM

Understanding human actions in video data plays an important role in computer vision researches. Owing to the widely use of bag-of-words approach, the size of bag-of-words is critical for performances of human action recognition tasks. However, the selection of the number of visual words mainly depends on empirical or subjective choices. In this work, we propose a data-driven approach to select the number of visual words via a minimization of the localized generalization error of Radial Basis Function Neural Network (RBFNN). Experimental results show that the proposed method yields good human action recognition accuracies for video datasets.