Development of intelligent wearables for the estimation of motion kinematics and kinetics

Research output: Book/ReportDissertations


Research units


Motion analysis gains more and more relevance because people are interested in gaining knowledge about their performance and it is possible to collect big data easily using wearable sensors. There are also more and more ways of collecting data in daily life. Nevertheless, motion analysis providing insight into joint kinematics and kinetics is still restricted to a laboratory set-up. To overcome this, the aim of this thesis was to take the first steps towards an easy-to-use and easy-to-interpret inertial-sensor-based motion analysis sys- tem. Due to the increasing amount of data available, the approach could be based on artificial intelligence models.
Artificial neural networks can be used to approximate the relationship between given inputs and outputs. During a training process, they learn to adapt their weights and biases to predict the output of unknown test samples. This thesis investigated if it is possible to predict the three dimensional angles and moments of the hip, knee and ankle joint and the ground reaction force using artificial neural networks based on inertial data only. For this purpose, a framework to simulate inertial sensors’ data based on marker trajectories collected by an optical system was developed and validated with a custom IMU system.
The results showed a good agreement of the predicted values and the ground truth data for gait and fast changes of direction. Employing data augmentation techniques to enlarge the dataset improved the results. The use of a fully-connected feedforward neural network resulted in a better pre- diction than the use of a recurrent LSTM neural network. Nevertheless, recurrent neural networks should still be considered for future work because they are able to make real-time predictions and do not need time normalised data like fully-connected feedforward networks. In a pilot investigation, the use of a convolutional neural network also seemed to be a promising approach. However, further steps are required to validate the methods developed in this thesis and bring an intelligent inertial-sensor-based motion analysis system towards application. The very promising results of this thesis prompt further research in this direction.
Original languageEnglish
Place of PublicationKöln
PublisherDeutsche Sporthochschule Köln
Number of pages60
Publication statusPublished - 2020

ID: 5426677


View graph of relations