Online Updating Appearance Generative Mixture Model for Meanshift Tracking

20
Online Updating Appearance Gen erative Mixture Model for Mean shift Tracking Jilin Tu, Thomas Huang Elec. and Comp. Engr. Dept. Univ. of Illinois at Urbana and Champaign Urbana, IL 61801 {jilintu, huang}@ifp.uiuc.edu Hai Tao Elec. Engr. Dept. Univ. of Calif. at Santa Cruz Santa Cruz, CA 12345 [email protected]

description

Online Updating Appearance Generative Mixture Model for Meanshift Tracking. Jilin Tu, Thomas Huang Elec. and Comp. Engr. Dept. Univ. of Illinois at Urbana and Champaign Urbana, IL 61801 {jilintu, huang}@ifp.uiuc.edu. Hai Tao Elec. Engr. Dept. Univ. of Calif. at Santa Cruz - PowerPoint PPT Presentation

Transcript of Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Page 1: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Jilin Tu, Thomas Huang

Elec. and Comp. Engr. Dept.Univ. of Illinois at Urbana and ChampaignUrbana, IL 61801{jilintu, huang}@ifp.uiuc.edu

Hai Tao

Elec. Engr. Dept.Univ. of Calif. at Santa CruzSanta Cruz, CA [email protected]

Page 2: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Content

• Problem Statement• Overview of tracking object with appearance variations• Overview of Meanshift Tracking • Meanshift tracking with online appearance updating• Generative model for histogram appearance updating• Experiments• Summary

Page 3: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Problem Statement

• Our Goal:– To track object of vast (but typical)

appearance variations by online appearance updating.

– To avoid drifting problem which is typical in most tracking algorithm with online model updating mechanism.

– To Infer object status by modeling the appearance variations.

Page 4: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Overview: tracking of object with appearance variations

• Subspace Model– Black1998

• Eigentracking: Model the object appearances by PCA.• Unable to incorporate novelties.

– Ross2004• Eigentracking with incremental PCA.• Handles appearance novelties better.

– Ho2004• Tracking with L-Infinity norm subspace.• Handles appearance novelties the best, but no theoretical gurranttee for avoiding drifting problem.

• Mixture model– Jepson2003

• Model the appearance variations as mixture of stable appearance, outliers and two frame variations caused by optical flow. • Drifting problem is not solved.

Page 5: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Overview of Meanshift Tracking– Meanshift(Chen1995):

• An iterative procedure for finding the data density modes (or cluster centers) of an ensemble of data samples.

• By shifting toward the mean of the samples in the vicinity.

Page 6: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Overview of Meanshift Tracking• Meanshift tracking

– Setting: The likelihood of the object appearance is represented by ensemble of likelihood pixel locations in a likelihood image. – Naïve Intuition (CAMSHIFT): Follow the cluster center using meanshift while the cluster center of the likelihood image moves.– Approach(Comaniciu2000): Maximizing the similarity of the histogram in the tracking window and the static histogram of the object appearance.

Page 7: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Overview of Meanshift Tracking• Meanshift tracking in a nut shell

Page 8: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Overview of Meanshift Tracking• Meanshift tracking properties

– Pros:• Realtime• Histogram appearance model tolarates minor appearance variations.

– Cons:• One single static histogram does not capture the appearance variations of the object.

Page 9: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Meanshift Tracking with online appearance updating• The framework

– The idea: updating the histogram with the constraint of key appearances.

Page 10: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Generative model for histogram updating

• The generative model

Page 11: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

EM rules for histogram updating

• Expectation

• Maximization

Page 12: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Online EM rules for histogram updating

• Expectation of Log-likelihood taking into the consideration of past observations

where• Online Maximization Rule

Page 13: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Histogram Updating Rule

• Given the generative model we have

)}({ * nh

}ˆ)}),({|({ wnzmp

)}({ nz

1w)}({ )1( nhL

)}({)},({ )2()1( nhnh LR

)}({ )2( nhR

2w

Page 14: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Experiments

• Application: Human head tracking with coarse head pose inference. – Head appearance changes substantially when

human turns his/her head. – Usually only near frontal head pose provides

more information about the person’s identity, facial expression, eye gaze, lip movement, etc.

– It is important to track the head and detect near frontal head pose simultaneously.

Page 15: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Experiments

Tracking with L-Infinity Subspace Updating

Page 16: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Experiments

Tracking by Meanshift Tracker

Page 17: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Experiments

Meanshift Tracking with online updating

Page 18: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Experiments

Meanshift Tracking with online updating: Another example

Pose estimation accuracy 77%

Page 19: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Summary

• Our tracker with online appearance updating– More robust comparing to the standard Meanshift tracker.– Better tracking accuracy.– Avoid drifting problem.– Can infer the object state in relation to the appearance variation. – Acquisition of more than one key appearance of the object sounds annoying, but is practical in most online/offline tracking tasks.

Page 20: Online Updating Appearance Generative Mixture Model for Meanshift Tracking

Questions