Human Action Recognition Based on Oriented Gradient Histogram of Slide Blocks on Spatio-Temporal Silhouette

Video can be regarded as three dimensional spatio-temporal volume, in which human action is a three dimensional shape (3D shape) surrounded by the spatio-temporal silhouette surface. The type of human action depends on the shape of the silhouette surface. In this paper, we proposed a new feature called Oriented Gradient Histogram of Slide Blocks by building dense overlapping spatio-temporal slide blocks to detect the shape of the 3D silhouette surface of the human action. Sparse coding is adopted to represent videos based on the new feature and Random Forest is utilized to classify the types of human actions. Experiments on KTH and Weizmann human action datasets demonstrate that the new feature can describe the spatio-temporal silhouette surface correctly, accordingly recognize the human action types accurately.