Research Of Human Indoor Motion Recognition Based On Multilayer Neural Network

With the rapid development of microelectronics,the performance of various sensors has been improved rapidly.More and more sensors are suitable for smart phones.Smart phones have become an indispensable part of life.The indoor positioning technology for smart phones is becoming more and more popular among scholars and researchers.The recognition of the movement status of personnel in indoor positioning is an important part of indoor positioning.Traditional personnel indoor motion status identifies the method of analyzing video data and wearable device data.Because this method is too expensive in all aspects,it becomes a hot topic to find a more adaptable method of indoor motion recognition.Indoor mobile state recognition for smart phones mainly consists of three steps: collecting data,building identification models,and classifying and predicting.A single sensor data type is used in the existing mobile status recognition model for smart phones,which can not effectively summarize the state characteristics of indoor motion,and the motion recognition model does not have the ability to learn repeatedly.In this paper,the above problems are further studied.Firstly,to improve the single sensor recognition model and analyze the indoor motion state data collected by multi sensors in the smartphone,the SVM,decision tree and KNN algorithm are constructed in the indoor motion state recognition model.Finally,a LSSVM recognition model is proposed on the basis of the SVM algorithm.The LSSVM algorithm model is used to reduce the time complexity of the SVM algorithm model.A good recognition rate of motion state is achieved.Secondly,in view of the poor self-learning ability in the basic learning algorithm and the problem of constantly learning to adjust the parameters of the model,a model of motion state recognition based on multi-layer neural network is proposed.The multi-layer neural network based on sparse self coding is an unsupervised network structure.The time domain information in the motion state data is retained by the introduction of window method.Through continuous training of the network,the parameters in the recognition model are corrected by the gradient descent method,and the higher recognition rate of the motion state of the learning algorithm model is obtained.Finally,aiming at the problem that the feature extraction needs too much human intervention in multi-layer neural networks,an indoor motion state recognition model based on depth learning is proposed.The typical CNN network in the field of deep learning is chosen as the adaptive network for the motion state recognition model.The CNN network has the feature self extraction,and it can extract the ability to express the characteristics of each motion state most.The local connection network structure can reduce the time complexity of the recognition model and improve the operation efficiency.The convolution and pool layer in the CNN network motion state recognition model can effectively avoid the local optimal in the training process,and effectively improve the stability and accuracy of the motion state recognition.