Simulation
Olympic Weightlifting Video Analyzer
A video analysis system for Olympic weightlifting that estimates ground reaction forces, barbell forces and mechanical power output from standard video footage, combining AI pose estimation with a flexible multibody barbell model and state-input estimation.
Links & Resources
The Gap
Commercial video analysis apps offer basic kinematic overlays for weightlifting. None provide the quantities that matter most for understanding performance: ground reaction forces (GRFs), forces applied to the barbell, accurate segmental kinematics, and mechanical power output. This system was built to fill that gap.
The Stack
Three methodological threads are combined in a single pipeline. AI-based computer vision — built around Raidyn's segmentation and tracking tools — handles barbell detection, pose estimation and motion extraction from video. A flexible multibody model of the barbell captures its elastic deformation under load, which is significant at competition weights and cannot simply be ignored. A combined state-input-parameter estimator then uses the multibody model's equations of motion to infer the forces applied to the bar from the estimated kinematics.
Once barbell forces are known, ground reaction forces can be estimated with considerably higher precision than methods based solely on video-derived accelerations. The model also generates a virtually unlimited supply of synthetic training data by forward-simulating the barbell under known loads, which enables robust AI segmentation training without requiring large annotated datasets of real lifting footage.
Validation on World-Class Footage
The system was tested on publicly available YouTube footage of Bulgarian lifter Karlos Nasar performing a record-breaking 229 kg clean and jerk at the European Championships in Moldova, at age 20. The footage was low resolution and at a lower frame rate than what would typically be used for in-house analysis — conditions considerably outside the training distribution. The algorithms extracted barbell kinematics, deformation, applied forces and power output without meaningful degradation.
Weightlifting consists of two movement patterns, which constrains the pose space significantly and helps the AI components generalize. Still, performance on unseen footage at degraded quality provides a useful lower bound on robustness.
Scope
The current system produces mechanical power output and force estimates. Joint-level force estimation is the logical next step, requiring segmental inertia properties and a full kinematic chain from barbell to ground contact.