Human Action Recognition Research Based On Kinect

With the improvement of computer vision technology application demand,the research of understanding human behavior becomes critical.While human action recognition has a broad application prospect in the field of computer vision,especially in the intelligent monitoring,human-computer interaction,virtual reality,motion analysis, etc.Therefore,human action recognition research is also attracting more and more attention. At the same time,recent introduction of cost-effective depth cameras brings on a new trend of research on body-movement action recognition.Currently,a lot of studies are based on the pre-segment video stream for human action recognition.And they can’t recognize human action online.Based on the research of human action recognition home and abroad,this paper uses spectral clustering and DTW algorithm to access the human skeleton data captured by the Kinect skeletal tracking technology and gets the structing skeletal feature based on the granularity of human body part movement.Then we get the action model using jointly sparse coding algorithm to select and classify the structing skeletal feature.Finally,we realize the online human action recognition. The following introduces the content of this paper in detail.First,download MSRC-12 Kinect Gesture Dataset and transform it into motion data stream according to human action representation to prepare for the generation of template dictionary and the feature extraction.Second,learn a template dictionary from the motion data stream above using spectral clustering algorithm of machine learning and improved Dynamic Time Warping(DTW) algorithm.Third,extract the structured skeleton feature of each frame in motion data stream,the feature of one frame is the distances of the best fitting subsequences ending at this frame and template dictionary.Fourth,learn a transformation matrix to transfer structured skeletal features into action labels using the jointly sparse coding algorithm and get a gesture model.Fifth,the skeletal data of each frame preprocessing is taken online to avoid noise interference. Sixth,translate the skeletal data to motion data stream using the method of the first step,and then extract feature of each frame from online motion data stream using the method of the third step.Seventh,transfer the feature of fifth step into action label using the transformation matrix of fourth step,then recognize the human action.There are two major challenges in this paper:i) how to continuously recognize actions from unsegmented streams,and ii) how to differentiate different styles of a same action from other types of actions.In this paper,we solve these two problems with a new effective and efficient feature extraction method that uses a dynamic matching approach to construct a structured skeletal feature vector for each frame and improves sensitivity to the features of actions within the same class.And then recognize human actions from online motion data streams.The innovation point of this paper : In the template dictionary generation and feature extraction phase,this paper adopts the improved DTW algorithm,thus increasing the number of the valid template and improving the efficiency of the online recognition.At the same time,in online recognition stage,skeletal data online preprocessing is taken in this paper to filter the uninterested and wrong skeletal data.Thus,it is effective to avoid the noise interference in online recognition.