Negative space template: A novel feature to describe activities in video

Abstract

In this paper, we propose an implicit silhouette based (template based) method of recognizing activities in video. The novelty of the proposed method is that it extracts features from the negative space to form action templates while other implicit methods employed positive spaces to construct action templates. Extracting features from negative space facilitates the system to extract simple, yet effective features to describe each action. These features are robust to deformed actions due to complex boundary variations, partial occlusions, non-rigid deformation and small shadow. Moreover, other template based methods need to apply additional processing to reduce the dimensionality of the extracted features, but negative space templates do not require dimensionality reduction since it describes actions with low-dimensional features. The proposed method improves its processing time while achieving comparable accuracies with other state of the art methods.