Human Action Recognition Method Based On Recurrence Plot And Convolutional Neural Network

In recent years,human action recognition technology has received widespread attention from researchers,and it has potential application value in many fields,such as medical health monitoring,interactive somatosensory games,and military affairs.Existing human action recognition methods based on machine vision are subject to interference from environmental factors such as light effects,object occlusion,and may infringe on users’ personal privacy.In recent years,with the continuous development of microelectronics and wireless communication technology,various mobile devices with independent computing capabilities and widely used in people’s lives have provided researchers with new ways to solve human action recognition problems.It has important significance in pervasive computing.Based on a mobile device such as a smartphone,this thesis collects three axis acceleration data during motion,and then analyzes and studies human action recognition methods.The specific work is as follows.Aiming at the source of experimental data,a data acquisition and preprocessing system based on acceleration sensors has been designed..Firstly,the types of human movements to be collected in this article were established through a questionnaire survey.Then,smart phones were used as experimental equipment to collect sensor data under various human actions.Finally,the triaxial acceleration data is preprocessed by filtering and denoising,and the experimental data used in this thesis is obtained.The classification feature dimension of human motion data obtained by time domain and frequency domain analysis is too high and there are nonlinear gradability problems.A human action recognition method based on support vector machine is proposed.Firstly,feature sets are extracted from experimental data in time domain and frequency domain.Then,principal component analysis is carried out on the feature set and further selection of classification features is completed to get the feature subset.Finally,support vector machine is used to recognize human action.The experimental results show that this method can achieve a high recognition accuracy for common human actions and prove that this method can be applied to human action recognition tasks.The manual feature extraction process in human action recognition tasks is complex and difficult to extract nonlinear features from temporal signals,a human action recognition method based on recurrence plot and convolutional neural network is proposed.Firstly,the original one-dimensional acceleration signal is reconstructed and preprocessed to obtain an RGB color image.Then,it is fed into a convolutional neural network to complete feature extraction and recognition of human actions.The experimental results indicate that this method is more suitable for human motion recognition tasks compared to existing studies,specifically in comparison to other studies using triaxial acceleration and triaxial angular velocity data,this method still has the best classification performance under the condition of only using triaxial acceleration data from the public dataset UCI HAR Dataset,with a classification accuracy of 98.37%.In addition,this method has also been validated on the UCF Dataset,which has complex types of human actions,with a classification accuracy of 98%.