A Review of Human Pose Estimation Methods in Markerless Motion Capture

Abstract

Human pose estimation aims at detecting human joint points from input data such as images and videos, or building a human body model for motion analysis. However, due to the ambiguities of human occlusion, depth blurring and the lack of training data, the accuracy of motion capture is still far from satisfactory. This paper reviews the advances of human pose estimation methods in markerless motion capture since 2019. We propose three types of representations for the human body, and detect that a unified volumetric model provides more detailed motion representation. We introduce datasets and evaluation metrics widely used for 2D and 3D pose estimation. Comparisons and discussions are conducted on different model frameworks for human pose estimation based on accuracy, robustness, and speed, summarizing the strengths and weaknesses of various methods. We discover that pose estimation methods based on the Transformer framework exhibit better accuracy and robustness, while kinematic and physical knowledge greatly assist in solving 3D pose estimation. Additionally, lightweight methods are often overlooked in research. In conclusion, this paper serves as a guide for researchers interested in the field and assists newcomers in selecting and developing human pose estimation methods.