Research On Human Action Recognition Method Based On Stereo Vision

Human action recognition is related to a lot of subjects, such as computer vision, machine learning, pattern classification, signal processing, image processing, artificial intelligence, and so on. It has significant theoretical value to study and also has large potential applications such as surveillance system, video analysis, intelligent robots, human-computer interaction and other fields. Because of the differences between the appearances of human body and the action habits, the inter-class variations are usually large while the extra-class variations may are small under different conditions. And at the same time, viewpoint changes, varying illumination conditions and occlusions can all alter the way actions are perceived. Therefore, it is very challenging and difficult to recognize human actions. In this thesis, daily indoor action recognition is studied based on the 3D skeletal data tracked by Kinect, and also an action recognition system is designed. The main contributions are as follows:1) The key steps in human action recognition are the extraction of action features and the selection of recognition algorithm. Kinect is used to obtain depth maps, and get the 3D locations of 20 human body joints with the help of released SDK. A posture representation method based on shape histogram is proposed. This method is scale invariant and translation invariant.2) In the study of recognition algorithm, because of that human action is a series of postures which can be considered as an observed state sequence, the Hidden Markov Model (HMM) is used to train and recognize. An action recognition algorithm based on HMM model is proposed.3) In order to complete the research work, an action recognition system is designed and implemented, and also an indoor action database containing 7 types of actions completed by different people in different viewpoint is established. A large number of experiments are performed on this database and public MSR-Action3D database. The experiments results show that algorithm achieves good recognition results, which indicated that the posture representation method and the action recognition method are feasible.