Development of biomechanical analysis-based human motion recognition algorithms

Human motion analysis and recognition that is a part of motion perception in computer vision has become an issue of strategic importance in recent years. Recognition systems developed for understanding of human behavior has been numerous research fields including criminal identification, security systems, virtual reality, computer vision and human-computer interactions and several proposals have been made in order to solve the recognition problems. Although various methods have been proposed for recognition of human activities obtaining different data from realistic videos, the dependencies and relations among human motions have not been much investigated. In the context of this thesis, the studies providing three main contributions have been presented. Firstly, a simulation method based on fuzzy logic with a novel adaptive method has been proposed for modeling and simulation of running motion. Simulation results of the method, which is applied to two-dimensional human model, have been tested. Secondly, biomechanical analysis of natural human motions such as sitting, standing up, jumping, walking have been performed and computer vision-based human motion recognition algorithms using this analysis have been developed. 3D data have been extracted from points of the moving joints with a sensor camera that can capture images at a certain speed, and it is constructed a simple and robust biomechanical model to recognition of human motions with these collected data. Finally, performed recognition methods with key-joint features and Principle Component Analysis (PCA) gains advantage using biomechanical information theory of human body as well as the practical requirements of near-real-time performance and accuracy. Experiments validate that the proposed methods outperform most existing methods and the model is computationally efficient. The results of the proposed methods are discussed and supported by several scientific publications.