Real-Time 3D Multi-Person Pose Estimation Using an Omnidirectional Camera and mmWave Radars
Aarti Amin, Alberto Tamajo, Isaac Klugman, Emil Stoev, Timothy Fisho, Hwasup Lim, Hansung Kim · 2023
Learning-based monocular 3D human pose estimation holds significant potential for a variety of applications, including sports, automation, and entertainment; however, not always at a cost that allows it to be scaled. This paper proposes an affordable solution to learning-based monocular 3D pose estimation from 2D videos that can be utilised outdoors and indoors. We introduce a system that leverages an omnidirectional camera and mmWave radars to estimate the 3D pose of the people in the scene in real-time. The proposed algorithm shows good pose reconstruction accuracy with the average Euclidean distance between a ground truth body joint position and its 3D reconstruction ranging from 4.5cm to 19cm within 20 meters along both the$x$and$z$axes of the camera.