Comparative Study of Adaptive Segmentation Techniques for Gesture Analysis in Unconstrained Environments

Abstract

This paper discusses the importance of providing non- invasive techniques in the analysis of the complex movements performed by musicians and athletes. Current gesture analysis systems are insufficient and do not succeed in providing quality performance measurements without imposing strict environmental and operational constraints on individuals. Computer vision offers the means with which such techniques can be made possible without impeding the performance of musicians or compromising the integrity of the measurements. The following study compares some of the modern image segmentation techniques and discusses their shortcomings with respect to the stated context. A novel statistical method is also introduced in an attempt to improve the resilience of vision-based gesture segmentation. I. INTRODUCTION With the advent of more powerful computing facilities and advances in artificial intelligence, few techniques have been proposed in the evaluation of the complex movements inherent in the performances of musicians and athletes. Such techniques are highly desirable in order to detect problematic situations which often lead to chronic injuries and to provide the capacity with which professors and trainers may measure the evolution and habits of individuals. Gesture analysis provides the means with which quantitative measurements may be obtained with respect to the physical performances of musicians and athletes. Modern gesture analysis methodologies employed by professionals fail to obtain an exhaustive evaluation or a complete comparison between gestures. Current techniques in this type of analysis are not robust enough to operate in the unconstrained environments in which these performances must be evaluated. Many techniques today still rely on the use of encumbering sensor technologies and require the use of cabling or attaching markers on an individual. The techniques also often require a very controlled environment and involve tedious, expensive and complex operating methodologies. The imposed environments are usually foreign to the musicians or athletes resulting in an impeded performance and corruption to the exactitude of the measurements. The undertaken study aims to research and develop new vision-based methodologies suitable for the analysis of musician gestures. Specifically the context being explored consists of analyzing the gestures and postures of pianists with applications to piano pedagogy. Computer vision techniques offer the ideal means for this type of evaluation. Due to their non-invasive nature they do not interfere with the performance of pianists. Computer vision techniques are put to the test in a context where the imposed constraints on the environment and individuals must remain minimal.