Human action recognition using normalized cone histogram features

In this paper, we propose a normalized cone histogram features method to recognize human actions in video clips. The cone features are extracted based not on the center of gravity as is common, but on the head position of the extracted human region. Initially, the head, hands and legs positions are determined. Thereafter, the distances and orientations between the head and the hands and legs are the extracted and employed as the features. The histogram's x-axis represents the orientations and the y-axis the distances. To make the method invariant to human region sizes, the features are normalized using the L2 normalization technique. The classification method used was the perceptron neural network. We conducted experiments using the ucf-sports-actions database to verify the effective ness of our approach. We achieved an accuracy of about 75% on a selected test set.