Using particle filter for face tracking

Particle filtersMaria Mikhisor

3D without glasses

Autosterescopicdisplay

Eye tracking

3D without glasses

Eye trackingTo simplify my task, I started from head tracking

My task is to track faces on different distances

There are many different algorithms for tracking…For various reasons I chose Particle Filters

Not this kind of particles…

A particle is a hypothesis, one possible state of a system

Prior distributionFrame 1

Prior distributionFrame 2

Prior distributionFrame 2

ObservationFrame 2

Averaging posterior

Frame 2

ResamplingFrame 2

Prior distributionFrame 3

MoveFrame 3

ResampleFrame 3

5 distances1m – 3m

3 videos foreach distance

3 experiments for each video

My experiments

Getting ground truth data:

1 video 1 min long labeling manually 1800 frames

Motion model

position 1st order

velocity

acceleration

2nd order3d order

Motion model

Observation models

Color histogram

Active shape model

Histogram of oriented edges

Haar wavelets

Combining feature detectors

Cascades

Linear combinations

Color histogram + Haar wavelets

Color histogram + Active contour

Simple histogram VS compound histogram

Ellipse parameterization

Number of particles

Plans for near future…

Cylindrical projection

Multi-cameras tracking

One 3D Particle FilterTwo 2D Particle Filtersthat communicate or

Use Kinect for collecting ground truth data