Markerless Motion Capture with Unsynchronized Moving Cameras N. HaslerN. Hasler, MPI Inf., Saarbrucken, Germany B. RosenhahnB. Rosenhahn, MPI Inf., Saarbrucken,

Markerless Motion Capture with Unsynchronized Moving

CamerasN. Hasler, MPI Inf., Saarbrucken, Germany

B. Rosenhahn, MPI Inf., Saarbrucken, GermanyT. Thormahlen, MPI Inf., Saarbrucken, Germany

M. Wand, MPI Inf., Saarbrucken, GermanyJ. Gall, MPI Inf., Saarbrucken, Germany

H.-P. Seidel, MPI Inf., Saarbrucken, GermanyIEEE 2009

Outline

• Introduction• Steps• Experiments

Introduction

Steps

• Camera Calibration• Camera Synchronization• Motion Capture

Camera Calibration

• Single Camera Structure from Motion KLT–Tracker or SIFT filter out feature points

RANSAC with multi-view constraints

minimizes the error of 3D points (Gaussian distribution)– d(….) = Euclidean distance– Pj = 3D object point– Ak = 3x4 camera matrix– p(j,k) = K images J trajectories of 2D feature point

Camera Calibration

• Multi camera Structure from Motion– Register N reconstructions into global system (H) find and merge 3D object points (pairwise match) (color intensity , uniqueness constraint)

Camera Calibration

• Tensor Voting filter (noise)• Least Squares filter (smooth)• 3D surface reconstruction

Camera Synchronization

• Camera Synchronization– Audio signals– denotes cross correlation (Fast Fourier Trans.)– * convolution

Motion Capture

• Kinematic Chains

• Silhouette Extraction

• Pose Estimation

Experiments