Assessment of the Accuracy of a Deep Learning Algorithm- and Video-based Motion Capture System in Estimating Snatch...

Area of Science:

• Biomechanics
• Sports Science
• Kinetics and Kinematics

Background:

• Quantitative kinematic analysis is crucial for evaluating weightlifting snatch performance.
• Marker-based (MB) systems are accurate but impractical for real-time training and competition.
• Markerless video-based (VB) systems offer a potential practical solution using deep learning.

Purpose of the Study:

• To assess the comparability and applicability of VB systems for obtaining snatch kinematics.
• To compare VB system kinematic data against a gold-standard MB reference system.

Main Methods:

• 21 weightlifters performed snatches at varying intensities (65-80% 1RM).
• Kinematics were captured using both MB (Vicon Nexus) and VB (Contemplas/Theia3D) systems.
• Analysis included joint center positions, joint angles, and Statistical Parametric Mapping (SPM).

Main Results:

• Significant differences were found in joint center positions (lower limb: 4.7±1.2 cm, upper limb: 5.7±1.5 cm).
• VB and MB systems showed varying agreement in joint angles, with better frontal plane correlation (RMSD: 11.2±5.9°).
• SPM revealed significant kinematic differences across most degrees of freedom, particularly in lower limb extension angles and velocities during the second pull.

Conclusions:

• Current VB systems exhibit significant kinematic differences compared to MB systems, indicating a lack of direct comparability.
• Differences likely stem from varied models and assumptions rather than inherent measurement accuracy.
• Advancements in neural network-based VB approaches show promise for future use in weightlifting analysis.