A Comparative Assessment of Accuracy in Video-Based Monocular Human Pose Estimation Frameworks

Area of Science:

• Computer Vision
• Biomechanical Analysis
• Machine Learning

Background:

• Comprehensive evaluation of human pose estimation frameworks is crucial for advancing research and practical applications.
• Focus on state-of-the-art 2D and 3D frameworks developed since 2019.

Purpose of the Study:

• To conduct a thorough review and comparative analysis of leading human pose estimation frameworks.
• To evaluate framework performance on a novel dataset using quantitative metrics.

Main Methods:

• Systematic review of 118 papers and 4 GitHub repositories.
• Selection of 16 frameworks including AlphaPose, Detectron2, MediaPipe, MeTRAbs, MHFormer, MMPose, MoveNet, OpenPifPaf, OpenPifPaf-vita, OpenPose, PoseFormerV2, rtmlib, StridedTransformer-Pose3D, ultralytics (YOLOv8), ViTPose, and YOLOv7.
• Evaluation on a custom dataset of exercise videos with synchronized motion capture data, using weighted mean absolute error and weighted intraclass correlation coefficient.

Main Results:

• MeTRAbs identified as the top-performing overall framework.
• AlphaPose, rtmlib, and YOLOv7 demonstrated superior performance in 2D human pose estimation.
• Quantitative analysis revealed significant differences in joint angle accuracy across frameworks.

Conclusions:

• MeTRAbs offers the best comprehensive performance for human pose estimation tasks.
• Specific frameworks like AlphaPose, rtmlib, and YOLOv7 are recommended for 2D applications.
• The study provides valuable insights for selecting appropriate frameworks for diverse human pose estimation challenges.