multi-camera tracking of articulated human motion using motion and shape cues

Multi-camera Tracking of Articulated Human Motion using Motion and Shape Cues

Aravind Sundaresan and Rama Chellappa

Center for Automation ResearchUniversity of Maryland, College Park MD

USA

What is motion capture?

Motion capture (Mocap) is the process of analysing and expressing human motion in mathematical terms. Initialisation, Pose estimation and Tracking.

Applications Motion Analysis for clinical studies, Human-

computer interaction, Computer animation.

Marker-based systems have shortcomings Cumbersome, introduce artefacts, time

consuming.

Marker-less system desirable.

Calibration and Human body model

Use multiple cameras (8) in our capture 640x480 grey scale images at 30 fps. Calibrated using algorithm of Svoboda.

Use articulated human body model. Super-quadrics for body segments. Model described by joint locations and super-

quadrics. Pose is described by joint angles.

Overview

Use images from multiple cameras. Compute 2-D pixel displacement between t and

t+1. Predict 3-D pose at t+1 using pixel displacement. Compute spatial energy function as function of

pose. Minimise energy function to obtain pose at t+1.

Tracking Framework

Use motion and spatial cues for tracking. Motion cues use texture.

Error accumulation: estimates only change in pose. Spatial cues obtained from silhouettes, edges,

etc. Instability: Solutions are stable only “locally”.

Predictor-Corrector framework. Predictor:

Compute motion(t) from pixel displacement. Predict pose(t+1) from pose(t) and motion(t).

Corrector: Assimilate spatial cues into single energy function. Correct pose(t+1) by minimising energy function.

Pixel registration and displacement

Project model onto image to obtain Body part label for pixel. 3-D location of pixel. Mask for each body part

Find dense pixel correspondence using Parametric optical flow-based algorithm for each

segment.. Minimise MSE:

Pose from pixel displacement

State-space formulation

Linearisation We show that Taylor series Iteratively estimate pose

Combine spatial cues

Combine multiple spatial cues into a single “spatial energy function”.Compute pose energy as function of dx, dy and Φ.

+ =

Minimise 3D pose energy

Given multiple views and 3-D pose Compute 2-D pose for ith image Compute Ei for ith camera using 2-D pose 3D pose energy, E = E

1+ E

2 + ... + E

n Compute minimum energy pose using

optimisation.

Tracking results