Human Action Recognition by Extracting Features from Negative Space

A region based technique is proposed here to recognize human actions where features are extracted from the surrounding regions of a human silhouette termed as negative space. Negative space has the ability to describe poses as good as the positive spaces (i.e. silhouette based methods) with the advantage of describing poses by simple shapes. Moreover, it can be combined with silhouette based methods to make an improved system in terms of accuracy and computational costs. Main contributions in this paper are two folded: proposed a method to isolate and discard long shadows from segmented binary images, and generalize the idea of negative space to work under viewpoint changes. The system consists of hierarchical processing of background segmentation, shadow elimination, speed calculation, region partitioning, shape based feature extraction and sequence matching by Dynamic Time Warping. The recognition accuracy of our system for Weizmann dataset is 100% and for KTH dataset is 95.49% which are comparable with state-of-the-art methods.