Event-Driven 3D Vision for Human Activity Analysis in Context of Dance and Fitness Training of Elderly People
Thesis by Thomas Hahn
Supervision by Hannes Kaufmann
Abstract
Over the last years many implementations concerning the recognition of human motion have been developed. In doing so different systems for human motion detection reaching from recognition of simple gestures to more dynamic complex motions have been invented. The application area of these systems is thereby wide spread from input for Human Computer Interaction to human motion analysis in the field of rehabilitation exercises or sports. Systems that are designed for elderly people are becoming more important, especially in the physical training application area. This is because the population is tending to live to an older age and there will be more and more elderly people in the near future. In this thesis a system for recognition of human motion in the area of dance and fitness training for elderly people is introduced. This module within the EU project Silvergame [SJSB09] is thereby intended to help elderly people to keep their level of health as well as to gain a higher fitness level so that they can stay healthy to an older age. With the system the users can then be encouraged to move more by performing the dance which they see on their home TV screen. In doing so such a dance consists of different human activities which the system recognizes. Furthermore, it also provides some sort of feedback via the given output device. As the input device, a novel event-driven 3D vision sensor, developed at the AIT Austrian Institute of Technology is used in this approach. What is special in this case is that only data is transferred if an intensity change in the field of view is detected. Therefore, less data then with ordinary video systems is generated. Another difference worth mentioning is that this information is communicated not frame-based but pixel wise. Keeping this constraint in mind and based on the information transferred from this sensor, elementary features that are used as input for classification are obtained. Through a detailed research of the literature about the up-to-date classification methods, the most promising technique and features for the motion detection system were chosen. This thesis thereby shows the performance of the designed application and points out the opportunity for further employments. Though it was significant how the chosen classification method can be used for the obtained features from the received data. Additionally first performance measurements were done. For this first implementation MATLAB was chosen as the main platform and further applications shall be based on this gained knowledge. For experimentation with the implemented algorithm a database including 580 samples with 8 different activities from 15 individuals, using the 3D sensor, was recorded. To obtain representative experimentation results a cross validation was applied and different settings were used to compare the results. Additionally, test sessions were done on different data sets and for the best results the training and evaluation time was recorded to point out the possibility of real-time usage. The best results thereby reached an average correct recognition rate of around 96 %.
Reference
T. Hahn: "Event-Driven 3D Vision for Human Activity Analysis in Context of Dance and Fitness Training of Elderly People"; Supervisor: H. Kaufmann; 188/2, 2011.
BibTeX
Click into the text area and press Ctrl+A/Ctrl+C or ⌘+A/⌘+C to copy the BibTeX into your clipboard… or download the BibTeX.