Automatic pose initialization of swimmers in videos

We propose an approach to the task of automatic pose initialization of swimmers in videos. Thus, our goal is to detect a swimmer inside a target video and assign an estimated position to her/his body parts. We first apply a non-skin-color filter to reduce the search space inside each target frame. We then match previously devised template sequences of Gaussian feature descriptors against sequences of feature vectors which are computed within the remaining image regions. Finally, relative average joint positions from annotated images featuring the key pose are assigned to the detection result and three-dimensional joint positions are estimated. We present detection results for test videos of three different swim strokes and examine the performance of four types of feature descriptors. 1. PROBLEM STATEMENT AND MOTIVATION For the past decade the analysis of human motion has been a heavily investigated subject by the computer vision community. Especially the analysis of sports videos has gained a remarkable amount of attention in computer vision and sports science research in recent years. One major goal is to automatically evaluate quantitatively an athlete’s performance under training and racing conditions continuously in order to help coaches to adjust the training schedule to an athlete’s personal needs as well as to analyze a competitor’s edge. A first key problem of continuous pose analysis is to find the location of an athlete inside of a video and to identify reliably her or his pose. This first key problem of pose analysis, usually referred to by pose initialization, is the subject of this paper. It commonly serves as the starting point for any subsequent pose tracking through time in a video. In this paper we concentrate on videos of swimmers for several reasons: On the one hand motion analysis is especially useful for individual sports such as swimming where the athlete’s success vitally depends on his actual movement patterns. On the other hand swimming videos feature two additional challenges compared to videos of most other sports: (1) An athlete is usually not fully visible from pool side cameras as it is hard to see below the water surface and (2) the background, which is the water surface of the swimming pool, is highly noisy, wavy and specular. Thus, in our framework we first identify and exclude background areas, before using feature descriptors to model the visible part of a swimmer’s body from a few template images. In general we can expect that a technique working under these difficult conditions should also be applicable to other visually easier observable individual sports such as running or long jump.