Research And Implementation Of Human Action Recognition Based On 3D Bone Data

Human action recognition is an emerging and challenging research field,which has great value in human-computer interaction,video surveillance,and smart medical fields.Traditional human action recognition is mostly based on 2D video and images,and the recognition effect is not good.With the development of 3D technology such as depth sensor,it can reduce the influence of illumination changes and environmental background in the recognition process.This method makes the human action recognition achieve better recognition effect.In this paper,the feature extraction method of 3D bone data is deeply studied,and a human action recognition method based on mixed local features is proposed.At the same time,the traditional VLAD algorithm is improved,and a human action recognition method based on improved VLAD is proposed.The effectiveness of this method is verified by experiments.Use this improved method to design and implement a human action recognition system.The main research work is as follows:1.Aiming at the shortcomings of the existing human action recognition method,the accuracy of expression is not high.A human action recognition method based on mixed local features is proposed.The method improves on the way of local feature extraction.,which extracts local features such as displacement vector,relative position and joint angle from bone joint points to construct mixed local features to describe different kinds of actions.This method not only describes the motion information of the action well,but also retains the information with strong discriminating ability and removes redundant information.2.The traditional VLAD algorithm only contains first-order mean information,and it is difficult to fully describe the distribution of actional features.An improved VLAD algorithm is proposed for this problem.The algorithm assigns each feature vector to multiple neighborhood center,which linearly combines these class center vectors under the criterion of least square error,and close to the corresponding feature vector.Meanwhile,the combined coefficient obtained in the previous step is taken as the membership degree,and the VLAD is calculated on the center of the plurality of categories with the membership degree as the weight.Finally,the algorithm aggregates the features into fixed-size descriptor vectors.This improved algorithm is combined with the human action recognition method proposed in this paper.It is verified by experiments that this method can better describe the distribution information of features.3.According to the human action recognition method proposed in this paper,a human action recognition system is designed and implemented.The system can read human bone data stored locally and identify current action.