A Novel Orientation-Context Descriptor and Locality-Preserving Fisher Discrimination Dictionary Learning for Action Recognition

Abstract

This paper presents a novel local posture orientation-context descriptor, and proposes a FDDL(Fisher discriminant dictionary learning) method based on local orientation-preserving(LOP-FDDL) for sparse coding in action recognition task. To take full use of the information about the position of the local body-part related to the center of the torso, ant the spatial-temporal shape changes of the human body-parts, we extract orientation-context descriptors of local body-parts to express the local posture of human body. Our descriptors not only include orientation information, but and also include the information of geometric structure and motion of body-parts. In order to accurately express action sequences, we need to learn a discriminative dictionary with strong expressive power which consists of the information about categories and orientations of body-parts from the extracted posture descriptors. Hence, a discriminative dictionary learning model based on the manifold constraint of local orientation-preserving is proposed, and Fisher Criteria is considered in the sparse coding stage of this model, which makes the coding coefficients discriminative. Meanwhile, to improve the performance of dictionary and learning efficiency, we initialize the dictionary as a class-structured dictionary which is a block-structured dictionary with orientation information. Experimental results demonstrate that our proposed method is better than other related action recognition methods on Weizmann and KTH public datasets.