Zhong Zhang

Graduation Semester and Year

2015

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

Department

Computer Science and Engineering

First Advisor

Vassilis Athitsos

Abstract

The broad application domain of the work presented in this thesis is human motion analysis with a focus on hand detection for American Sign Language recognition and fall detection for assistive environments.One of the motivations of the proposed thesis is a semi-automatic vision based American Sign Language recognition system. This system allows a user to submit as query a video of the sign of interest, or simply perform the sign in front of a camera. The system then asks the user to annotate the hands' locations in the sign. Next, the hand trajectory of the query sign is compared with the models in a large sign database to find the best matches. At last, the user reviews the top results to verify which of them best matches the query sign. Towards making the system more automatic, a novel hand detection method is introduced which is a combination of four representative hand detection methods published in these years.On the topic of fall detection for assistive environments, the work in this thesis aims at improving the safety of patients and elderly persons living unaccompanied at home. More specifically, this thesis proposes a fully automatic vision based fall detection method which can serve as a component of a home monitoring system for elderly people. The major contributions of the fall detection work can be summarized as: (i) This thesis collects three kinds of fall datasets using Microsoft Kinect depth cameras: non-occlusion dataset, partial occlusion dataset and complete occlusion dataset. The non-occlusion dataset refers to the performer being always visible to the camera when he/she falls down. A partial occlusion fall refers to a fall where part of the body is occluded by a certain object when the person performs the fall action. When the end of a fall is totally occluded by a certain object, like a bed, the fall is called a complete occlusion case. All of these datasets are freely available online, together with annotations marking the beginning and end of each fall event. As far as we know, this is the first public fall datasets captured by depth camera. These datasets will enable researchers to explore their own fall detection methods. (ii) This thesis proposes a statistical fall detection method based on a single Kinect depth camera, that makes a decision based on information about how the human moved during the last few frames. Our method proposes novel features to be used for fall detection, and combines those features using a Bayesian framework. The proposed method is quantitatively compared with three most related publications which also use a single depth camera on the collected datasets. Experimental results demonstrate that the proposed method obtains much better detection accuracy than other competitors on non-occlusion and partial occlusion datasets. As for the complete occlusion dataset, although the proposed method does not get the best detection accuracy, the evaluation between the proposed method and the competitors can be taken as a benchmark for the assessment of more advanced fall detection method.

Disciplines

Computer Sciences | Physical Sciences and Mathematics

Comments

Degree granted by The University of Texas at Arlington

Download

Share

COinS