Research On Human Action Recognition Based On Multivariate Time Series

Human Action Recognition(HAR)refers to the recognition of human actions from information such as time series or images through methods.Human action recognition is the basis of many fields such as human-computer interaction,security monitoring,motion recording,behavior research,etc.It has important research meaning and broad development prospects.At present,with the development of Micro-Electro-Mechanical technology,wearable devices are gradually increasing.These devices or programs record the acceleration information of human actions in the form of multivariate time series.How to quickly and accurately identify actions from data that have large volumes and lack labels is a challenging problem.This thesis conducts experiments on the existing methods based on the PAMAP dataset,and the experimental results show that the current methods are inferior.This thesis proposes a new method,Distribute-Based State Extract(DBSE).After many experiments,the experimental results show that the proposed method has higher accuracy and lower time consumption.The specific research content of this thesis includes the following parts:(1)Use Toeplitz Inverse Covariance-based Clustering(TICC)to extract actions from multivariate time series.The TICC method is a representative method for multivariate time series segmentation and clustering,and its results are representative.The experimental results show that the TICC method has an inferior ability to extract actions,and the time consumption is excessive.This method is not suitable for analyzing human actions recorded by multivariate time series.(2)Research on the statistical similarity of acceleration sequences of multiple actions,it is found that similar actions have strong statistical similarity,and there is no similarity between different actions.It can be seen that statistics are suitable for describing actions.Design experiments to verify the validity of the selected statistics.The experimental results show that the statistics selected from the probability density distribution can describe acceleration sequences.(3)The thesis proposes Distribute-Based State Extract(DBSE)method.Use the PAMAP dataset for multiple experiments.Experiments show that this method can quickly and clearly cluster the actions from the multivariate time series.Compared with other clustering algorithms,it is found that the clustering performance of the DBSE method is improved by 17%,the clustering Purity reaches 0.967,and the time consuming is greatly reduced.In addition,a quantitative study on the size selection of sub-sequences was conducted.Experiments were conducted on the time series irregular and uneven sampling that may occur when multiple sensors work together.Experimental results show that the DBSE method has better robustness in two situations.(4)Combine the DBSE method with the K-Nearest Neighbor(KNN)algorithm to realize the recognition of actions.Use Ball-Tree to optimize KNN's nearest neighbor search.Experiments show that this method can achieve a 93.7% action recognition rate when the labeling rate is insufficient,which is better than backpropagation neural network(BP),support vector machine(SVM)and other classification and recognition methods.