Comparison of Grid-Based Dense Representations for Action Recognition

Abstract

Dense descriptors have recently been widely used for human action recognition. Especially several methods of grid-based dense representations have been proposed in the literature. In order to evaluate which method performs better in the field of action recognition in unconstrained environments, four kinds of grid-based dense descriptions and recognition methods are tested on UCF sports dataset in this paper. And the recognition results of histogram of optical flow (HOF), histogram of oriented gradient (HOG), pyramid histogram of oriented gradients (PHOG) and Gist descriptor are compared. Furthermore, this paper also compares and analyzes recognition effects using 1-Nearest Neighbor (1NN) and support vector machine (SVM) combining these descriptors. It shows that the combination of Gist descriptor and SVM to reach the best recognition accuracy on UCF sports dataset.