Large Area 3D Human Pose Detection Via Stereo Reconstruction in Panoramic Cameras

We propose a novel 3D human pose detector using two panoramic cameras. We show that transforming fisheye perspectives to rectilinear views allows a direct application of two-dimensional deep-learning pose estimation methods, without the explicit need for a costly re-training step to compensate for fisheye image distortions. By utilizing panoramic cameras, our method is capable of accurately estimating human poses over a large field of view. This renders our method suitable for ergonomic analyses and other pose based assessments.