Sparse representation-based human action category recognition using an action-context region-aware dictionary

Abstract

Human action recognition is a core functionality of systems for video surveillance, content-based video indexing, video search, robotics, and human-object interaction. The goal of vision-based human action recog-nition is to determine the action label that best describes the action instance in the video clip automatically. Conventional vision-based systems for human action recognition require the use of subject segmentation tech-nique in order to achieve an acceptable level of recognition effectiveness. However, generic techniques for au-tomatic segmentation that guarantee high performance are currently not available yet. Further, it is important for effective human action recognition to use the features associated with a context region adaptively, as well as localizing the actions. To resolve the problems, this paper proposes a novel sparse representation-based method for human action recognition, which does not require complex human or human action localization for test video clip. To that end, we construct a dictionary for SR-based classification of human actions that con-sists of two split dictionaries: an action region dictionary and a context region dictionary. The action region dictionary aims at characterizing the action regions in a non-segmented test video clip (such as human action region, saliency region), while the context region dictionary aims at characterizing the non- action regions in the test video clip under consideration. That way, we are able to segment, implicitly, action region and context information in a test video clip, thus improving the effectiveness of classification. That way, we are also able to develop a context-adaptive human action recognition strategy. As shown by comparative experimental re-sults obtained for the UCF-50 and UCF-sports Action data set, the proposed method facilitates effective human action recognition, even when testing does not rely on explicit segmentation.