This repository contains the simulation code and mathematical proofs for OccluPose, a framework designed to recover joint trajectories during severe visual occlusion in athletic environments.
The core of this project is a spatiotemporal inference engine that uses Gaussian Process (GP) Regression to bridge gaps in computer vision data. Instead of relying on purely spatial heatmaps (which fail when a limb is hidden), OccluPose treats movement as a continuous probabilistic function bounded by Newtonian physics.
Standard Human Pose Estimation (HPE) models like HRNet or OpenPose treat every video frame as a discrete event. When an athlete's limb is blocked by a hurdle or equipment, these models "lose" the joint, leading to jitter or physically impossible coordinate jumps (limb collapse).
OccluPose solves this by formulating the recovery as a Maximum A Posteriori (MAP) optimization problem, using temporal momentum to maintain kinematic fidelity where visual data is zero.
Due to the proprietary nature of high-end athletic tracking data, this repository utilizes a Biomechanical Proxy Simulation to validate the math.
- Trajectory: A 3D Newtonian reconstruction of a right-knee hurdle jump (60 FPS).
- Occlusion: A 20-frame "blind window" simulating a total loss of visual signal.
- Baseline: An HRNet-W32 architecture simulation with spatial noise diffusion.
This script requires numpy, matplotlib, and scikit-learn.
# Clone the repository
git clone [https://github.com/Dev-Scodes5/OccluPose-Simulation.git](https://github.com/Dev-Scodes5/OccluPose-Simulation.git)
# Run the inference simulation
python occlupose_proxy.py