Human Body Part Classification in Monocular Soccer Images

This paper addresses the problem of finding body parts in images, which can be an essential first step for body pose estimation. The core component of the presented method is the pixel-based classification of body parts using Random Forests. This technique is applied to find the body part positions of soccer players in broadcast images. As this approach is usually used with depth data, we analyze how this method can be adapted to work with monocular images. In this context we identify the image representations with the best classification results. Although monocular images leave some ambiguities, our approach to body part classification achieves satisfying results: 90.32% of the pixels in the test set are correctly classified.