Human Action Segmentation And Recognition Based On 3D Skeleton Data

Human action segmentation and recognition is an important part of human action analysis.Although a lot of research has been done,human action segmentation and recognition still have challenges.For example,most unsupervised learning based hu-man action segmentation methods exist the phenomenon of excessive segmentation,and the intraclass variation and interclass similarity increase the difficulty of human action recognition,etc.Focusing on the problem of segmentation and recognition of human action,this paper firstly introduces the current research status of human action segmentation al-gorithm based on boundary detection and clustering and human action recognition al-gorithm based on traditional machine learning and neural network learning.Based on this,this paper proposes a human action segmentation and recognition algorithm based on 3D skeleton data,and the specific work is as follows:(1)Aiming at the problem of human action segmentation,this paper proposes a human action segmentation algorithm based on sequence labeling,which labels the action frames by means of supervised learning.This paper firstly extracts the spa-tial features of human action,including body joints locations,skeleton angles and the movement direction of the body joints.Then,the Bidirectional Local-Global Long-short Term Memory Network is proposed to extract the local-global temporal features of human action.Next,action frames are labeled by Conditional Random Field which is fused with Intersection over Union loss factor.Finally,human action is segmented by analyzing the labeled frames,which is the output of the sequence labeling model.The performance of this algorithm is verified on the UTKinect dataset,and this algorithm achieves high precision and recall rate.(2)Aiming at the problem of human action recognition,a human action recog-nition algorithm based on feature fusion is proposed in this paper,which analyze the characteristics of human action from different aspects.Firstly,each action is decom-posed into multiple subactions,and each subaction is analyzed by multiscale wavelet transform,which extracts the time-frequency features of each subaction.Based on this,the global feature of human action is obtained by combining the time-frequency features of subactions organically.At the same time,Relation-aware Long-short Ter-m Memory Network is proposed in this paper to extract the spatial-temporal features of human action by learning the relationships of human body parts in time sequence.Then,the time-frequency features and spatial-temporal features are fused into a more efficient feature by Auto-Encoder.Finally,the category of human action is output by softmax regression model.The performance of this algorithm is evaluated on three public datasets,including NTU-RGB+D,UT-Kinect and Florence 3D actions.The experimental results on these large datasets show that the accuracy is improved.