Human Activity Recognition And Embedded Application Based On Convolutional Neural Network

In recent years,with the rapid development of sensor technology,human motion recognition based on sensor technology has been more used in various fields.How to choose an appropriate classification algorithm pair to improve the accuracy of activity recognition is the current research hotspot.At present,classification algorithms are generally divided into two categories: activity recognition algorithm based on traditional machine learning and activity recognition algorithm based on deep learning.Generally,after preprocessing activity data,classification algorithm is used to extract data features to realize activity classification.However,there are not many public data sets of activity recognition,and the experimental environment is difficult to fully match.When practical application is needed,it is often necessary to collect data independently and establish a data set in line with the experimental environment.On the other hand,the current research on activity recognition mostly stays in the theoretical stage,and the research on the practical application of the model is not much.Therefore,this thesis also deploys the designed model to the embedded system based on ARM to complete the embedded application of human activity recognition.The main contents of this thesis are as follows.Firstly,in order to solve the problem of incomplete matching between the open data set and the experimental environment in this thesis,12 testers are selected to collect the X,Y,Z three-axis acceleration data of six kinds of human activities by using MPU6050 acceleration sensor.After a brief visual analysis of the original data,the original data is filtered by the sliding mean filter algorithm,and then the sliding window segmentation algorithm is used to divide the data into appropriate sizes according to the activity cycle.Finally,the feature extraction of the data is completed,and the data set required for the experiment is established.Secondly,after using the feature data to build the activity recognition model based on SVM,RF,GBDT,LR four traditional machine learning algorithms and using the original data to build the model based on LSTM network and CNN two deep learning algorithms,the visual comparative analysis of the six models is completed.Based on the final training effect of the model,the CNN model with high accuracy is selected to deploy to the ARM embedded device.And the practical application of human motion recognition is completed in the ARM embedded terminal.Thirdly,this thesis takes STM32H743 as the core processor to complete the hardware and software design of the embedded end.Because the recognition algorithm is limited by the memory resources of the current wearable devices,it is difficult to deploy to the wearable devices.This thesis uses STM32 CUBE-AI toolkit to deploy the convolutional neural network model to the active recognition devices,so that it can run on the ARM embedded system.Finally,through the experimental verification,the overall recognition rate of CNN model on the embedded device is 84%,and the average recognition rate of standing and sitting still is 90%.