Vision-based human pose estimation for sports biomechanics

Abstract

In computer vision, the process of estimating the body configuration in a single, monocular image is known as human pose estimation. Due to the human body articulation, the task of the human body model extraction is challenging and demanding. Therefore the problem of human pose estimation has been studied over the last two decades and has been evolving rapidly. The construction of human body models is achieved utilising different approaches. Some of these methods are upon the classical use of hand-crafted features, and some are based on deep learning and convolutional neural networks. The importance of human pose estimation lies on the abundance of applications, such as assisted living, sports analysis, and medical applications. The objective of the dissertation is to investigate the performance and the reliability of vision-based human pose estimation approaches that are markerless and based on convolutional neural networks. Specifically, the OpenPose human pose estimation approach is investigated in the context of sports biomechanics and rehabilitation to determine where this approach provides reliable and accurate results and where it fails. The OpenPose approach is considered one of the powerful state-of-the-art models for human pose estimation, which employs convolutional neural networks for both feature extraction and body parts and pairs detection. It is the first approach for 2D multi-person pose estimation using the Part Affinity Field for body part association. The investigation process is carried out by testing the OpenPose approach on the publicly available Leeds Sports Pose dataset, which contains a total of 2,000 images. The results obtained indicate that the overall performance of OpenPose is outstanding with high accuracy and reliability rates for body parts detection. However, this performance depends on the type of poses and the location of the individual parts in a given image. The OpenPose model is more reliable in detecting upper body parts than detecting the lower body parts. It particularly struggles in detecting the right/left ankle, but provides an outstanding detection rate for right/left shoulders. Also, it is challenging for the OpenPose to detect people in an upside-down position, such as athletes performing parkour, where in most cases it fails entirely. In addition, OpenPose is able to detect persons in an image if and only if either the nose or the neck keypoint is not occluded, otherwise the whole COCO skeleton is lost or the pose is not detected.