Research On Activity Recognition Method Based On Wearable Devices

With the popularity of micro-sensors(accelerometers,magnetometers,pressure gauges,etc.)on smart devices,many new research fields have been created,among which activity recognition based on wearable devices is an important application.Activity recognition is commonly used in security,surveillance,and entertainment,such as gait authentication,smart homes,fall detection for the elderly,and VR games.At present,most of the researches on activity recognition based on wearable devices have problems such as poor discrimination between similar activity,uneven sample distribution,and difficulty in feature extraction,which affects overall performance.This thesis addresses the issues mentioned above and solves them through the following aspects;Firstly,an activity recognition system was designed and built.The FXOS8700CQ sensor chip produced by Freescale(NXP)was used as a data acquisition terminal,and corresponding APKs were developed for data storage and upload,The Tomcat server finally implements the "Terminal-Client-Server" trinity activity recognition system data acquisition platform for subsequent research.Secondly,the collection of similarity movement datasets was completed,which contains four confusing movements(walking,jogging,stairs,and stairs).A total of 10 subjects without movement disorders were collected and each person performed once with the guidance of staff at the designated location.Finally we collected more than 50,000 original data.In addition,in order to verify the generalization of the algorithm,we also use two common datasets and perform corresponding preprocessing on the datasets.Thirdly,to solve the problem of low recognition rate of non-equalized data in similarity activity,Synthetic Minority Over-sampling Technique(SMOTE)+ simple random sampling(Resample)technology is used to classify common datasets,and then Then using five different algorithms to analyze it and synthesize different performance indexes.It is concluded that Random Forest(RF)and Multilayer Perceptron(MP)have the best classification effect in the similarity activity recognition,and the average accuracy rate reaches 99%,and have good robustness to sampling frequency and unbalanced data.Fourthly,to improve the difficult problem of feature extraction,the convolutional neural network model built in TensorFlow is directly applied to the original time series data in the self-made dataset.This eliminates the tedious steps of extracting the feature value and finally achieves a recognition rate of 96.67%after the parameters are optimized.In addition,we also extracted 43 eigenvalues from the self-made dataset,compared the two algorithms(RF,MP)that performed better under the feature extraction method with the experimental results of this scheme,and once again verified the superiority of this scheme.To sum up,this thesis has designed and built an activity recognition system,and proposed corresponding solutions of the problem of unbalanced dataset and low recognition rate of similarity activity,which improves the recognition rate of activity.In addition,the convolutional neural network is applied directly to the original time series data,this can eliminate the tedious steps to extract feature values,thus simplifying the entire process.