| dc.description.abstract | Inertial Measurement Units hold a great potential for real-time gait phase measurements, gait
kinematics measurement, and gait analysis. They offer convenience in the form of their light weight
and portability. There are however known drawbacks in the form of sensor drift and susceptibility
to noise.
Extrapolating data for gait analysis directly from IMU’s can therefore be problematic, but emerging
Machine learning / Artificial intelligence (ML/AI) modalities holds the potential of estimating
gait characteristics to a new level of accuracy. Applications such as interpolation of user intent in
prosthetic device control could through this be improved further, as well as movement analysis, and
gait identification (GI) applications.
In this thesis, IMU sensor data is used in combination with machine learning (ML) models to
extrapolate significant information about the users walking gait. The information is extrapolated
implicitly, through Machine Learning algorithms. The is centered around two applications: Gait
detection (GD, i.e Gait phase detection), and Gait Identification (GI, i.e identify human subjects
based on their walking gait). Both applications is already present in the scientific literature. The
computational load for some of the systems is however quite high, and not fit for a live system
running on a microcontroller or a mobile device. The proposed solution attempts to eliminate
some of the computational load by reducing the dimension of input data, and reducing ML model
complexity.
The work yields two implementations of Machine Learning on a Gait Identity task, and includes
discussion and analysis of important features for a ML Gait detection task for a dataset consisting
of 18 features.
For the first Gait Identity task, a TCN ML model architecture scored a perfect 1.0 on a Max
F1-Averaged Stratified k-Fold validation for k=5 on a self-acquired single-IMU dataset with 13
subjects.
On the second Gait Identity task, a TCN ML model architecture scored a 0.98 on a Max F1-Averaged
Stratified k-Fold validation for k=5 on a publicly available single-IMU dataset with 30 subject | |
