Dimensionality reduction and clustering in Gait Analysis
Hintsanen, Antti (2018-02-21)
Dimensionality reduction and clustering in Gait Analysis
Hintsanen, Antti
(21.02.2018)
Tätä artikkelia/julkaisua ei ole tallennettu UTUPubiin. Julkaisun tiedoissa voi kuitenkin olla linkki toisaalle tallennettuun artikkeliin / julkaisuun.
Turun yliopisto
Tiivistelmä
The walking pattern of a human is fairly unique for each individual. The pattern of locomotion is known as gait and the systematic research of gait is called gait analysis. Gait analysis not only makes it possible to distinguish individual people, but it can also help to find certain medical conditions that affect the gait. Gait is relatively similar for each healthy person, but patients suffering of medical conditions including Parkinson's disease (PD), stroke, and cerebral palsy often share similar abnormalities in gait.
In addition to separating healthy people from each other, detecting of these diseases is possible using gait. Continuous gait analysis can also help to detect some of these medical condition already at an early stage and help to measure and plan rehabilitation when recovering of such conditions.
Several methods for gait detection exist, although many of them are complex and expensive. Optoelectronics, footswitches and pressure sensors are among the instruments used in modern day gait analysis. Development of wearable sensors has opened new possibilities in gait analysis.
This thesis proposes a complete system for collecting gait data and analyzing it using machine learning algorithms. The wearable data collection system is based on a single six degrees of freedom (6DoF) sensor combining an accelerometer and a gyroscope. The algorithm for classification is based on principal component analysis (PCA) and K-Means clustering.
The main challenge for this study was finding a suitable method for efficient classification of persons by the gait data from only a single wearable sensor. Splitting the walking data into individual steps set another challenge. Thirdly a design for a complete wearable system is proposed to answer the question of how to make this system portable.
Suitability of the chosen algorithm was verified using a set of gait data collected using a wearable system built at University of Turku. The combination of PCA and K-means clustering proved to be effective in classifying the people. The clusters of the steps corresponded to the actual persons of the steps with an accuracy of 89.7%.
In addition to separating healthy people from each other, detecting of these diseases is possible using gait. Continuous gait analysis can also help to detect some of these medical condition already at an early stage and help to measure and plan rehabilitation when recovering of such conditions.
Several methods for gait detection exist, although many of them are complex and expensive. Optoelectronics, footswitches and pressure sensors are among the instruments used in modern day gait analysis. Development of wearable sensors has opened new possibilities in gait analysis.
This thesis proposes a complete system for collecting gait data and analyzing it using machine learning algorithms. The wearable data collection system is based on a single six degrees of freedom (6DoF) sensor combining an accelerometer and a gyroscope. The algorithm for classification is based on principal component analysis (PCA) and K-Means clustering.
The main challenge for this study was finding a suitable method for efficient classification of persons by the gait data from only a single wearable sensor. Splitting the walking data into individual steps set another challenge. Thirdly a design for a complete wearable system is proposed to answer the question of how to make this system portable.
Suitability of the chosen algorithm was verified using a set of gait data collected using a wearable system built at University of Turku. The combination of PCA and K-means clustering proved to be effective in classifying the people. The clusters of the steps corresponded to the actual persons of the steps with an accuracy of 89.7%.