Abstract des Autors

Underwater video analysis is common in elite level swimming for stroke correction, technique analysis and as a visual aid to support coaches and athletes. Because of the challenges of the underwater visual environment and ambulatory camera work, video can be degraded and considered non ideal, additionally the athletes movement can lead to blurred vision and bubbles from cavitation effects. In this paper we use image analysis techniques to enhance images for improved clarity and automated detection of limb segments related to metrics of interest at the elite level. Thus we address the problems of detection, segmentation and estimation of body configuration using several image processing algorithms. These classic imageprocessing problems are even more complicated with video footage of poor quality as described above. Our approach is to use adaptive algorithms like the global probability for boundary detection to detect the swimmer’s boundaries
and deformable models arranged in a pictorial structure to recognize body and limbs plus their configuration. Our results show that it is possible to detect the athlete and relevant body segments. Identification and estimation of the body configuration in every frame is demonstrated to be feasible. From this analysis, common metrics can also be extracted such as stroke counts, stroke rate and primary phases of an arm stroke. The goal of this work is to support the training of athletes and help them improve their technique.