Presenter

Miss Xiao Lan Yao [Email]

Abstract

The motor skills of a human being will inevitably decline over the human life span. Analyzing this slow decline and developing prediction models for associated fall and mobility risks have received considerable research attention lately. An accurate assessment will increase the quality of life, decrease the cost for the healthcare system and benefit to the sustainable development of the entire human being on the long-term. However, accurate assessment of the loss of motor skills requires specific tests to be performed on a very regular basis. The instrumented Timed Up and Go (iTUG) test is a classical test to estimate the mobility and the status of the balance of people by performing a series of consecutive activities of daily life (ADLs). Specifically, to accomplish these activities, the participant is asked to sit in an armed chair first, then rise, stand, walk 3 meters forward, turn, walk 3 meters back to the armed chair, and sit down again. If the total time of the test exceeds a given threshold, this is considered a clear sign of loss of mobility and an increase in fall risk. In our work, for the first time, a 3D camera is used to fully automatically analyze the iTUG test, by measuring the durations of each component of the test.
The first part of the thesis focuses on the estimation of the extrinsic parameters of the 3D camera. This is a crucial step before developing the entire automatic iTUG test system. Compared to traditional approaches for camera calibration, we propose two extrinsic camera calibration methods for 3D cameras without using any calibration objects with known geometric properties and without measuring the camera position in the world coordinate system. We validated the two methods both on simulated data and on real data. The experimental results show an acceptable accuracy of 1cm approximately. Furthermore, state of the art approaches for the accelerometer and gyroscope calibration are also described in detail and validation experiments are presented.
Inspired by the work of Wall J.C et al, in the second part of my work, the entire iTUG test is segmented automatically in 4 time intervals: sit-to-stand, walk forward, walk back, and turn-to-sit again. On the basis of analyzing a primary dataset including 17 healthy adults’ iTUG video sequences, an automated iTUG segmentation system using Dynamic Time Warping (DTW) shows an excellent correlation between the automated segmented time intervals and the manually segmented ones. On average, the correlation coefficient of the entire iTUG test is 88% and the MER (mean error rate) is 9%.
ii
Further validation of the proposed automated iTUG system is performed on a database collected from 16 volunteers of 63 years old. The correlations between the automated and manual segmentation points reach above 95%. The results on the elderly subjects hence confirm the previous conclusions. A comparison between the young and the aged group shows the differences both on the entire iTUG test and on the segmented sub-durations.
Some directions for future work are introduced at the end of the text: for instance Hidden Markov Model (HMM) based human activity segmentation, a method based on detecting and tracking the chessboard pattern attached to the subject, and a wearable sensor based method are all possible extensions.

Short CV

Master of Engineering, China University of Petroleum (East of China), 2009