Research On Human Action Recognition Based On 3D Skeleton Data

Human action recognition is one of the important research contents in the field of computer vision,It can replace our eyes to understand the human action intention,With the rapid attack of artificial intelligence technology in recent years,it has been more and more favored by researchers.Human action recognition has significant applications in many fields,such as medical monitoring,public security,human-computer interaction,and game entertainment.The existing methods can be roughly divided into the following categories according to the types of data processed by the computer:RGB data,RGB-D data,2D and 3D skeleton point data of human body.RGB data carries a large amount of redundant information,so in addition to the disadvantages of large amount of calculation,it is also susceptible to interference from factors such as illumination,background change and occlusion,which have a great impact on the recognition accuracy and speed.Therefore,this paper adopts 3d skeleton point data based on depth information as the feature of human action recognition.In view of the current problem that the accuracy of the action recognition algorithm is not high enough and the actual action recognition system is difficult to realize,the research content of this paper mainly includes the following aspects:1.First,in this paper,we propose a feature which is obtained by cooperative warp of two discriminative features.The two features are respectively skeleton point location features and speed features after pretreatment.We use Dynamic Time Warping(DTW)algorithm to extract them from the skeleton sequence.In feature extraction,we use the method of cooperative warp and we call it the feature of Cooperative Warp of Two Discriminative Features(CWTDF).In the following experimental part,the recognition accuracy of this feature proposed by us exceeds that of many existing methods in the three benchmark datasets such as MSR-Action3D,Florence3D-Action and UTKinect-Action datasets.In order to prove the superiority of the CWTDF feature,we also consider another scenario,which also use the two features above.However,when using the DTW algorithm for feature extraction,the two features are extracted separately and then fused.We call it the feature of Separate Warp of Two Discriminative Features(S WTDF).In the following experiments,it is also proved that the recognition accuracy of this feature is inferior to that of CWTDF feature in the three benchmark datasets.This conclusion also shows that the feature obtained by cooperative warp using DTW algorithm is more capable of representing an action than that obtained by separate warp,which is also one of the contributions in this paper.2.In order to further improve the accuracy of action recognition,this paper proposes an algorithm that uses machine learning to automatically extract useful information from original features while the improvement in accuracy is very limited by using traditional feature extraction methods.The name of our algorithm is action recognition algorithm based on subgroup feature learning.According to the natural structure of human body,the algorithm divides the skeleton joins of human body into five parts,namely,human limbs and trunk.Each part contains three skeleton points.Then we calculate the relevant features of each part respectively.This division is based on the fact that when we are doing a certain action,it is usually a combination of one or more parts of the body's limbs and trunk.In the following experimental part,it is also proved that this method can provide enough information for subsequent feature learning,so that the classification accuracy can reach a relatively high level.Then we design a loss function in the feature learning stage.Because we finally adopt the K-Nearest Neighbor classification algorithm,the purpose of the loss function is to make the Euclidean distance between the same category features as small as possible and the distance bet,ween different category features as large as possible after the original feature X go through the linear transformation(left multiplication transformation matrix L).After continuous learning iteration,the recognition accuracy of the algorithm is improved.After the tests on three benchmark datasets,the accuracy of our action recognition algorithm exceeds that of many papers.We also use the experimental method to select the relevant parameters in the algorithm to achieve a relative high level classification accuracy in our algorithm.3.This paper also designs a action recognition system based on Kinect.The system can use the Kinect camera to recognize human actions in real time.We have written the system in C#.The system uses a total of 12 angles of the upper and lower bodv of the human body as the features of each frame,and uses the method of average extracting frames to obtain 12 frames from one action sequence to form a vector as the feature representation of this action.The classifier is a 3-layer neural network,which we wrote in python.After training the model in python,we copy the weight matrix into C#.The data set is collected by our own Kinect device.After testing,the classification accuracy of our system is 85%.Although the accuracy of this method is not as high as the two methods above,the advantage of this method is that the feature calculation is very easy,and it is easy to realize in the actual system(C#),and the angle features used by us is not affected by the scale of the skeleton.