Continuous Human Action Recognition Based On Kinect

Action recognition is an important part of the field of computer vision.In 1960 s,a psychologist had studied on the human behavior and set the basic theory that trajectory can be used for action recognition.Over the years,the traditional action recognition technology has one or more defects in identification of efficiency,cost,and environmental constraints,which makes the application of this technology relatively limited.And the interval of complex actions,temporal and spatial distribution led the continuous action recognition task extremely difficult.With the appearance of low cost depth cameras such as Kinect,more and more researchers have realized that action recognition based on depth image and skeleton information can change the awkward situation.However,many of these studies are based on segmented or manually tagged action sequences,unmarked continuous action sequence identification remains less.This paper focuses on raising the correct rate of continuous action recognition algorithm,then respectively studies the classification of feature based on depth data and joint data and segmentation algorithm based on sliding window,in order to achieve the ultimate purpose that identifying continuous action.The main contributions of the study are as follows:1)An action recognition algorithm based on the depth feature and the feature of joint points is studied.For the balance of recognition accuracy and algorithm complexity of continuous action,an identification algorithm based on hybrid structure is used.In this algorithm,two kinds of weak classifiers using low complexity feature are connected in parallel,and then the parallel structure is connected with a strong binary classifier in series to form a strong multiple classifier with low complexity,witch can identify actions accurately and simply.The experimental results show that the recognition accuracy of the proposed algorithm is significantly improved compared with that of the weak classifier2)The strong binary classifier has a great influence on the accuracy of continuous action recognition in the hybrid structure.So an algorithm based on conditional probability criterion is studied.Screening joints which are useless for classification,which will eliminate all joints which make conditional probability of two kinds of models approximate,finally to improve the recognition accuracy and reduce thecomplexity of model.3)A scheme of motion segmentation and recognition based on SVM classification confidence is studied and verified.The algorithm use voting strategy of multiple classifier and the principle that functional margin can predict the confidence level of classification.This paper presents a method for calculating the confidence level of multi-classification SVM.With that,we use sliding window to find the peak for getting the starting point and action type.In the experiment,the algorithm is tested and analyzed based on the MSR Action3 D general action data set,and the correctness and effectiveness of the algorithm is verified.