Abstract:
Monitoring gait characteristics is an important tool used in many areas including orthopedics, sports, rehabilitation and neurology. Current methods applied to analyze the gait need clinical settings and equipments for measuring gait parameters. In this study, we propose an unobtrusive and comfortable system to perform gait analysis. Smartwatches equipped with embedded sensors including accelerometer and gyroscope are used to extract three main parameters of gait: step length, swing time and stance time. Data is collected from 26 healthy and volunteer participants with di erent ages and genders in clinical settings. Subjects wore smartwatches on both wrists, data is collected from two sensors: accelerometer and gyroscope. The data is preprocessed and step features are extracted. Relevant gait parameters are estimated using various regression models and compared with the ground truth data coming from the clinician using the golden standard instrumented walkway. Four machine learning algorithms including Linear Regression (LR), Gaussian Process Regression (GPR), Support Vector Machine (SVM) and Regression Tree, and two neural network architectures Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) are used to t data. Performance of the models is measured using a basic error metric, i.e. RMSE. The best model tting the data is found as GPR. Its RMSE value for the step length (cm) estimation is calculated as 5.29 cm. Besides the placement of sensors is less convenient than the state of the art studies, the gait analysis with smartwatches gives promising results and encourages for extended future studies.
