Date of Award
8-1-2013
Document Type
Masters Thesis
Degree Name
M.S.
Organizational Unit
Daniel Felix Ritchie School of Engineering and Computer Science
First Advisor
Jun Zhang, Ph.D.
Second Advisor
Bradley Davidson
Third Advisor
Mohammad H. Mahoor
Keywords
Bed-exiting detection, Bed-exiting prediction, Fall detection, Kinect sensor, MHI, Vision-based
Abstract
Fall is one of the most dangerous and costly accidents that threaten health of elderly people, and a large portion of falls occurs when a patient is trying to exit a bed. This thesis proposes two vision-based approaches for general fall detection and bed-exiting detection for elderly people, respectively. The Kinect sensor is chosen as the major monitoring device.
The first approach exploits the Kinect sensor with its Windows SDK to detect fall activities. The recorded spatial coordinates of the human body joints from Kinect's 3D skeletal view are processed to extract posture features. Then the principle component analysis and k-means clustering algorithms are applied for dimensionality reduction, vector quantization and feature translation. HMMs are well known for their application in temporal pattern recognition, thus they are chosen for this project to classify human motion which is a temporal sequence of postures. HMMs are trained by the labelled extracted features to model and discriminate four fall motion classes and three non-fall classes.
The second approach utilizes segmented motion history image (MHI) sequences to extract space-temporal features of a moving human body. Eight Hu image moments are calculated to translate the space-temporal features of each frame into vectors to describe video frames. The k-means clustering and HMM modelling are utilized for vector quantization and classification between bed-exiting activities and rolling-on-bed activities. In addition, likelihood probability curves are generated along the time line of all MHIs, endeavoring to predict a bed-exiting activity.
Detailed descriptions of the experiments and result evaluation are documented in this thesis. The experimental results using human subjects verifies the feasibility and effectiveness of the proposed approaches for general fall detection and bed-exiting prediction.
Publication Statement
Copyright is held by the author. User is responsible for all copyright compliance.
Rights Holder
Xiaoxiao Dai
Provenance
Received from ProQuest
File Format
application/pdf
Language
en
File Size
64 p.
Recommended Citation
Dai, Xiaoxiao, "Vision-Based 3D Human Motion Analysis for Fall Detection and Bed-Exiting" (2013). Electronic Theses and Dissertations. 152.
https://digitalcommons.du.edu/etd/152
Copyright date
2013
Discipline
Electrical engineering