Research On Human Motion Recognition Based On Local Error Model And Convolutional Neural Network

With the popularization and development of intelligent sensor devices such as smart phones and smart bracelets,the human activity recognition based on wearable sensors has become an important research hotspot in universal and mobile computing.It plays an increasingly important role in the fields of human behavior monitoring,health monitoring,medical rehabilitation,intelligent home scene and so on.In recent years,with the development of deep neural network technology,more and more researchers have begun to use convolutional neural network to solve the problem of human activity recognition.The recognition algorithm based on convolutional neural network can automatically extract human activity features and achieve better classification performance.However,deep convolutional neural network usually has many layers,large parametersand more memory,while the memory space of wearable devices based on embedded is limited.In the traditional cross entropy error training mode,all parameters of the hidden layer must be stored in memory until the forward and reverse error propagation ends.As a result,memory used to store hidden layer parameters cannot be released and reused,and memory utilization efficiency is low.The backward locking phenomenon limits the deployment and implementation of deep convolutional neural network on wearable sensor devices.Based on this,the paper proposes a convolutional neural network model using local error,which is used for the recognition of human activity based on sensors.Compared with the traditional global error,the local error proposed in this paper can train convolutional neural network layer by layer.Each layer of parameters can be trained independently according to local error and does not depend on the gradient propagation of the upper and lower layers.Thus,the memory used to store all hidden layer parameters can be released in advance,and it is unnecessary to wait until the end of forward and backward propagation,avoid the backward locking problem,and improve the memory utilization efficiency of convolutional neural network deployed on embedded wearable devices.The main contents of this paper are as follows:1.The human activity dataset are preprocessed by using sliding window technology,including designing suitable sliding window and sliding step,and normalizing and centralizing the processed sensor data.2.This paper constructs a local error training function for a single hidden layer by combining traditional cross entropy loss and similarity matching loss,and builds a convolutional neural network for human activity sensor tasks.On this basis,the optimal weights of traditional cross entropy loss and similarity matching loss are studied.By comparing the performance of the five public data sets of OPPORTUNITY,WISDM,UNIMIB-SHAR,PAMAP2 and UCI-HAR with the benchmark convolutional neural network model using global error,it is verified that the convolutional neural network model with local error has higher performance classification accuracy,lower memory usage.3.This paper further builds a deep residual neural network.Aiming at the characteristics of the deep residual neural network with jumpers structures,by combining the traditional cross-entropy loss and similarity matching loss,a local error training function for each residual module unit is designed.And compared with the benchmark residual network on the above five public data sets,a large performance improvement was obtained,indicating that the algorithm has good scalability.This paper compares the classification performance between the cross-entropy loss,the similarity matching loss and their combination through further ablation experiments.The results show that the single similarity matching loss is better than the cross-entropy loss,and the combination of the two has the highest classification performance.In this paper,the confusion matrix is further calculated,and the reliability and practicability of the algorithm are verified by deploying the model on smart phones and embedded devices for performance evaluation.