Improved 2D Human Pose Tracking Using Optical Flow Analysis

Abstract

In this paper, we propose a novel human body pose refinement method that relies on an existing single-frame pose detector and uses an optical flow algorithm in order to increase quality of output trajectories. First, a pose estimation algorithm such as OpenPose is applied and the error of keypoint position measurement is calculated. Then, the velocity of each keypoint in frame coordinate space is estimated by an optical flow algorithm, and results are merged through a Kalman filter. The resulting trajectories for a set of experimental videos were calculated and evaluated by metrics, which showed a positive impact of optical flow velocity estimations. Our algorithm may be used as a preliminary step to further joint trajectory processing, such as action recognition.