Research And Application Of Human Activity Recognition Based On Inertial Sensor

As a result of continuous advances in manufacturing processes,inertial sensors are becoming more and more compact and portable,rendering it possible to wear devices with them for periods.In recent years,applications such as rehabilitation assistance systems,somatosensory games,sports assistance systems,etc.that respond by sensing human behavior have become more and more common,thus making it urgent to study human activity recognition activities.Current activity recognition algorithms suffer from low performance and insufficient accuracy.For this reason,this paper focuses on how to improve activity recognition algorithms and validate them on three publicly available datasets,PAMAP2,UCI and WISDM,and finally implements an activity recognition application aiming to promote the research and development of activity recognition,with the following main work:At the beginning of this paper,the activity characteristics and the characteristics of the data after inertial sensor acquisition are analyzed,and a sliding window is selected for its data division,with a window length selected as 128 time points and a sliding length of 50 time points.For its signal characteristics,outlier cleaning is performed,and then a 3rd order Butterworth low-pass filter is used with a cutoff frequency set to 20.Finally,the three data sets are subjected to a data balancing process,which means that the number of each action is similar.Secondly,machine learning-based activity recognition is studied in this paper,and four models,Random Forest,Ada Boost,XGBoost and Light GBM,are selected.Eleven dimensionless features,11 quantized features,two frequency domain features and Pearson correlation coefficients were selected as training features according to the characteristics of the signals.The improvement over the same type of models was 16.5% and 2.48% on PAMAP2 and WISDM.The top 10 features of the four algorithms on three datasets were also analyzedSince the effect of activity recognition based on machine learning is still to be improved and feature extraction is very troublesome,another activity recognition algorithm based on deep learning is studied in this paper,and the whole model is built with the help of Bi-GRU,residual blocks and attention mechanism.The model is comprehensively examined in terms of model structure,model layers and accuracy,and the accuracy on PAMAP2,UCI and WISDM is improved by 11.51%,1.297%and 3.25% respectively,which proves the effectiveness of the model.Finally,an activity recognition application is designed and built in this paper,with the main modules of data acquisition and activity monitoring.In the data acquisition part,two ways are implemented,including receiving the external inertial sensor data by Bluetooth and acquiring the built-in sensor data of the cell phone directly.In the activity monitoring module,all activities that occur during the activity are identified by the activity recognition algorithm and displayed on the screen in chronological order.Quantization is used to compress the deployed model,improving the model speed by two thirds.