Online Human Action Recognition Research Based On Probabilistic Characteristics

The emergence of the depth camera, has greatly promoted the development of human action recognition. The depth image captured by depth camera can provide the depth information of the target, which can make the researchers highly efficiently and reliably extract the human skeleton data. Because skeleton data can describe the dynamic characteristics of the action, there have been many methods of human motion recognition based on skeletal data in recent years, which shows a very good recognition results. But in the field of human action recognition still exist some problems have not been effectively addressed:(1) how to continuously recognize action from the continuous action streams;(2) how to solve the intra-class variability of action;(3) how to improve the real-time of human action recognition method.To solve problems existing in the field of human action recognition, this paper presents a method of online human action recognition based on K inect. Is divided into the following three contents:(1) we use human skeleton representation based on Lie group to represent each moment of the human skeleton. Compare with other skeleton representation, the method based on Lie group can better describe the dynamic characteristics of action;(2) we propose a method that we extract the probabilistic characteristics by constructing a probabilistic model dictionary. We take an action as a fusion of movements of different human body parts, so we can use a small amount of body movement style to express a large amount of action style. Based on this opinion, we cluster the single dimensional motion time series of each individual part of the human body. After the comp letion of the cluster, the model of each class is modeled by the Markov chain, and the probability model of the variation law of the time series is obtained. The probabilistic models for all classes form the probabilistic model dictionary. Probabilistic model dictionary stores the movement of different styles of different parts of the human body. Then we will continuous action streams of each frame and a number of frames as an action instance and calculation the action instance in each dimension motion time sequence belongs to probability model dictionary corresponding dimensional probability model probability, the probability that all values of the final probability feature vector is composed.(3) Based on the combination of the feature extraction method and the Joint sparse coding algorithm, we proposed a new method of online human action recognition. Joint sparse coding is a linear multi class classification algorithm, which can effectively realize the fast online identification of continuous motion data.In this paper, the proposed method will test in MSRC-12 action data set for verification. The results showed that out method is about 3% higher than the methods at present in MSRC-12 data sets, and with a very high application value with respect to the method in this paper.