Geometry-informed correction of projection bias in browser-based monocular squat assessment

Area of Science:

• Biomechanics
• Computer Vision
• Human Movement Analysis

Background:

• Monocular RGB-based human pose estimation is valuable for field-based movement assessments.
• Existing systems suffer from uncharacterized projection-related errors compared to lab standards.
• Simplified camera geometry in markerless motion capture (MMC) introduces systematic bias.

Purpose of the Study:

• To quantify projection bias in a browser-based MMC system.
• To validate a geometry-informed linear correction framework for sagittal-plane squat analysis.
• To assess the accuracy of corrected joint angles under standardized monocular acquisition.

Main Methods:

• Simultaneously recorded bodyweight squats using 3D marker-based optical motion capture (OMC) and a 2D webcam.
• Developed and validated a geometry-constrained linear correction model using leave-one-subject-out cross-validation.
• Analyzed sagittal-plane squat kinematics under standardized monocular conditions.

Main Results:

• Raw MMC joint angles showed significant underestimation (hip: -11.2°, knee: -10.6°).
• The validated correction model effectively neutralized systematic bias (hip: 0°, knee: 0°).
• Root mean square error was substantially reduced, and coefficients of determination significantly improved for hip and knee joints.

Conclusions:

• Projection-consistent geometric factors are primary error sources under controlled camera conditions.
• The geometry-informed correction framework mitigates systematic errors in monocular pose estimation.
• Corrected joint angles offer practically relevant estimates for strength assessment and rehabilitation monitoring without anthropometric data.