Geometric Algebra Representation And Ensemble Action Classification Method For 3D Skeleton Orientation Data

Human body action classification and recognition based on 3D skeleton data has received more and more in-depth exploration and research attention due to its wide applications in entertainment,monitoring,human-computer interaction and other fields.However,the existing feature extraction methods is performed on the entire skeleton data of an action,so that the action representation needs to wait for the end of the action.This leads to several drawbacks.For example,the computation amount is large in the classification process and the classification delay is long.In addition,due to the different perform speed of the same action and the length of the different actions,the number of frames in the extracted skeleton feature are different.So additional representation encoding methods are needed to transform the extracted action representations to obtain a uniform dimension to satisfy the requirements of the input dimension of the subsequent classification system.To this end,this paper proposes a single-frame human pose orientation feature descriptor based on 3D skeleton data under geometric algebra framework and an ensemble human motion classification and recognition method that can be performed online in real time.The steps are as follows:1.Geometric algebra combining geometry and algebra,is independent of the coordinate system and can intuitively describe the rigid body in space and calculate the rotation efficiently.In view of the fact that the human body is composed of bones that can be regarded as rigid bodies,which are hinged by joints in space.The Euclidean space's bone data is mapped to the geometric algebraic space,to extract the single-frame human body orientation feature descriptor,which is composed of the human body's most informative bone orientation and the angle with respect to the trunk bone.It has strong explanatory power and high computational efficiency,and its classification and discrimination ability can better characterize the human body posture.2.After extracting the single-frame human body orientation feature descriptor,each frame of the human motion data set is used as an input to train the single-frame human body posture classifier.In the process of classification and recognition,the action sequences to be classified are successively predicted by the trained classifier to obtain a classification result sequences.Meanwhile,the integration method is adopted to integrate the classification result sequences.Finally,the action with the largest number of categories is taken as the classification result.The integrated human motion classification method based on single-frame human posture data has the advantages of simple flow,high classification accuracy,and can be used for real-time online action classification and recognition.In this paper,the public data set SYSU-3D-HOI and the skeleton data set SZU-3D-SOEARD collected during the study were respectively used for single-frame classification and motion classification experiments.The experimental results show that the proposed human gesture orientation feature descriptor has a good discrimination ability for human posture,and the proposed integrated human motion classification and recognition method has the characteristics of simple flow,high classification accuracy and strong realtimeness,which can be used to carry out the online classification task of human motion in real time.