Wrong Pose

Deep learning framework for 3D human pose quality assessment and ergonomics scoring from video streams

Collaborative project developed for the Deep Learning & CNNs course at the University of Technology Sydney (UTS). Code and documentation available on GitHub.

Workplace musculoskeletal disorders (MSDs) and repetitive strain injuries are strongly correlated with prolonged poor posture. Wrong Pose is an automated computer vision service that processes continuous video streams to detect body posture, reconstruct 3D joint configurations, and score ergonomic correctness in real time.

Real-time keypoint extraction and frame-wise pose assessment.

Technical Pipeline

  1. Person Detection & 2D Keypoint Regression: Employs Detectron2 (Faster R-CNN with ResNeXt-101 backbone) to localize persons and regress 17 body keypoints per frame.
  2. Temporal 3D Kinematics: Passes 2D keypoint trajectories into VideoPose3D (strided 1D dilated convolutions) to resolve depth ambiguities and reconstruct full 3D body motion.
  3. Ergonomic Quality Regressor: Computes 8 joint angle metrics and feeds scale-invariant feature vectors into a calibrated Multi-Layer Perceptron (MLP) regressor achieving 98% test accuracy against ground-truth ergonomics guidelines.
Ergonomic posture score breakdown across monitored joint angles.