Links
Abstract
A common approach in current research publication on analysing multimodal time-series data from inertial sensors is to use Convolutional Neural Networks (CNNs) to automatically learn patterns in the given data. The disadvantage when using CNNs in such an application on mobile devices is that they are highly computationally complex. Therefore, this thesis studies the question if Long-Short-Term-Memory (LSTM) networks, which are significantly less computationally complex, can be used as an alternative for such applications. Firstly the topic of Human Activity Recognition (HAR) and the used datasets Opportunity and mHealth are presented. Following that, inertial sensors and their functional principles are briefly touched on. In the main part of the work, the methods used, mainly LSTMs and CNNs are laid out. In our experiments, we mainly present an architecture study in which 380 LSTM variants were trained and tested on the Opportunity dataset. A modified training process called Relaxed Truth Training that speeds up the training process slightly is also presented and tested in our experiments. When comparing the best two LSTM networks from the architecture study to two different CNN architectures, we found superior performance from the LSTMs, mainly when looking at close to real life, continuous classification problems. These results were verified by testing the same architectures on the mHealth dataset.
Bibtex
@mastersthesis{Jabs2020RecurrentNeuralNet, author = {Jabs, Christoph}, school = {Reutlingen University}, title = {A Recurrent Neural Net Approach to Activity Recognition}, year = {2020}, thesis_type = {B. Eng. thesis}, }