Thesis
Signal information processing tools for healthcare diagnostics
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2020
- Thesis identifier
- T15603
- Person Identifier (Local)
- 201459256
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- The smart healthcare monitoring service has been given more attention in recent decades. With rising healthcare demand and progress in image processing, video-based gait assessment becomes a good alternative solution to assess the physical recovery progress for post-stroke survivors. However, most video-based assessment systems, commercially and in the literature, usually requires large laboratory space, are of high cost, and not portable, thus are impractical for in-home use. Accurate, low-cost, portable motion capture systems are growing in popularity, especially those that do not require expert knowledge to operate. This research proposes an alternative single depth camera based OPTIcal Kinematics Analysis system (named 'OPTIKA'). Novel signal processing and computer vision algorithms are proposed to determine motion patterns of interest from infrared and depth data, and enable real-time simultaneous tracking of joints based on attached retroreflective ball markers. Specifically, an accurate trajectory-based gait phase classification system is proposed to facilitate the diagnostics of muscle activities during gait, using readings from low-cost motion capture systems 'OPTIKA'. Feature selection/extraction methods are proposed to enable an automatic segmentation of motion records into individual gait cycles with nine gait phases slice, which provides a more intuitive diagnostics experience for clinical therapists to analyze the rehabilitation progress associated to the kinematics in particular gait periods. This research also analyzes the sensitivity of feature selection/extraction methods against the classification performance in two healthcare monitoring applications. To overcome the limitations of high-cost training data labeling work and when parts of the training labels are noisy, a robust semi-supervised binary classifier is proposed to combine deep learning and graph based signal processing methods. The experiments demonstrate that given an acceptable proportion of noisy training labels, the proposed classifier outperforms several state-of-the-art classifiers. The overall concepts and systems presented in this thesis form an underlying approach for further video-based healthcare monitoring service that assists the diagnostics of physical rehabilitation.
- Advisor / supervisor
- Stanković, Vladimir
- Stanković, Lina
- Resource Type
- DOI
- Date Created
- 2020
- Former identifier
- 9912922589702996
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T15603 | 2021-07-02 | Public | Download |