Detection of postural transitions using machine learning
Tutor / directorCatalà Mallofré, Andreu
Document typeTutored research work
Rights accessOpen Access
The purpose of this project is to study the nature of human activity recognition and prepare a dataset from volunteers doing various activities which can be used for constructing the various parts of a machine learning model which is used to identify each volunteers posture transitions accurately. This report presents the problem definition, equipment used, previous work in this area of human activity recognition and the resolution of the problem along with results. Also this report sheds light on the process and the steps taken to undertake this endeavour of human activity recognition such as building of a dataset, pre-processing the data by applying filters and various windowing length techniques, splitting the data into training and testing data, performance of feature selection and feature extraction and finally selecting the model for training and testing which provides maximum accuracy and least misclassification rates. The tools used for this project includes a laptop equipped with MATLAB and EXCEL and MEDIA PLAYER CLASSIC respectively which have been used for data processing, model training and feature selection and Labelling respectively. The data has been collected using an Inertial Measurement Unit contains 3 tri-axial Accelerometers, 1 Gyroscope, 1 Magnetometer and 1 Pressure sensor. For this project only the Accelerometers, Gyroscope and the Pressure sensor is used. The sensor is made by the members of the lab named ‘The Technical Research Centre for Dependency Care and Autonomous Living (CETpD) at the UPC-ETSEIB campus. The results obtained have been satisfactory, and the objectives set have been fulfilled. There is room for possible improvements through expanding the scope of the project such as detection of chronic disorders or providing posture based statistics to the end user or even just achieving a higher rate of sensitivity of transitions of posture by using better features and increasing the dataset size by increasing the number of volunteers.