Skeletal Graph Based Human Pose Estimation in Real-Time

Human pose estimation is a vivid topic in current literature due to its wide-spread applications such as motion-capture, telepresence or object manipulation in virtual environments. The process of human pose estimation is concerned with finding the pose parameters of a human body model that best fit to the observations in one or more input images. There exists a variety of algorithms that solve this task with high accuracy from multiple input images, depth images or even a single photograph. Unfortunately, these systems often require manual initialization and cannot process images at interactive frame rates. Often this can be overcome by learning poses from thousands of examples or fitting a rigid body part model to the data. However, model fitting algorithms are easily distracted by missing or spurious body-parts and depend on a good initialization. Therefore there is need for improvement in real-time body pose estimation methods to handle the full articulation space of the human body, support automatic single-frame initialization and tolerate outliers.