Using Synthesized IMU Data to Train a Long-Short Term Memory-based Neural Network for Unobtrusive Gait Analysis with a Sparse Sensor Setup.

In this work, we evaluated the possibility to use synthesized IMU data for training a deep neural network to generate a more complex, full-body description of the human gait in terms of joint angle trajectories from a sparse sensor setup. In this context, a sparse sensor setup consists of a few sensors attached to human body segments in an unobtrusive manner to possibly provide a monitoring system in an everyday life scenario. Since the relation between the input IMU data and the output joint angle trajectories is highly non-linear, neural networks appear to provide an optimal framework to formulate a mapping description. Especially with respect to periodic signals, recurrent neural networks (RNNs) have gained importance in the recent years. In this work, we have used a special type of RNNs that can be implemented by using long-short term memory (LSTM) cells, which have shown promising results when being applied to sequential data. The artificial training data was generated by a simulative human gait model and virtually attached sensor devices. The trained network was subsequently validated by a dataset that was recorded from a treadmill walking trial using a motion capturing system and an IMU sensor system. The qualitative comparison already shows promising results, however, this study can only be considered to provide preliminary results in this area. Clinical Relevance- This approach has the potential to be applied in the remote assessment of gait behavior during everyday life environments using an unobtrusive sensor net-work. In particular for monitoring older people suffering from an increased fall risk or any significant gait impairments this work is of possible interest.