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

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human Motion Analysis

Background:

• Unobtrusive human gait monitoring is crucial for everyday health assessment.
• Sparse sensor setups offer potential for non-invasive gait analysis.
• Accurate joint angle trajectory estimation is vital for understanding gait dynamics.

Purpose of the Study:

• To evaluate the use of synthesized Inertial Measurement Unit (IMU) data for training deep neural networks.
• To develop a method for generating complex, full-body human gait descriptions from sparse sensor data.
• To explore the application of Recurrent Neural Networks (RNNs), specifically Long-Short Term Memory (LSTM) cells, for gait analysis.

Main Methods:

• Generated artificial training data using a simulated human gait model and virtual sensors.
• Employed Long-Short Term Memory (LSTM) recurrent neural networks to map IMU data to joint angle trajectories.
• Validated the trained network using data from a treadmill walking trial with motion capture and IMU systems.

Main Results:

• The trained deep neural network demonstrated promising qualitative results in generating joint angle trajectories from IMU data.
• The study established the feasibility of using synthesized data for training gait analysis models.
• Preliminary findings suggest the potential for accurate gait description from sparse, unobtrusive sensors.

Conclusions:

• Synthesized IMU data can effectively train deep neural networks for human gait analysis.
• LSTM networks show promise for modeling the non-linear relationship between IMU data and joint kinematics.
• This approach has potential clinical relevance for remote gait assessment and monitoring individuals at risk of falls.