Model base human pose tracking
Papers Real-Time Human Pose Tracking from Range Data Simultaneous Shape and Pose Adaption of Articulated Models using Linear Optimization Both are model-based: use a given mesh to estimation pose that matches input frame.
Real-Time Human Pose Tracking from Range Data Key point: Extend ICP based framework to incorporate so called “free space” constraint
Model-based Simple Mesh Model: each part as a 3D capsules Dynamic Bayesian Network
Formulation
Pipeline Three steps: – Holding x constant, maximize the objective with respect to correspondences c – Nearest Point Search – Update Joint Locations: gradient ascent – Correct Joint Locations using Constraints
Constraints Bone Length – the two affected joint positions can be moved in 3D space along the capsule center line until the constraint is met
Constraints Free space – Constraining all points on the model surface to lie in the 3D region of space behind the measurements – Approximated by two separate constraints Silhouette – If a model point projects outside the silhouette, we project it to the closest point inside. Z-Surface – If a model point projects outside the silhouette, we project it to the closest point inside.
Results See paper
Paper 2 Key point – Proposed “differential bone coordinates” to bine the surface with bone for joint optimization
Basic Pipeline Iterate between – Find correspondence: covariance weighted nearest neighbor – Minimize
Covariance Weighted NN Each pair is weighted by the mean of covariance of the two vertices
Energy Minimization