Date of Award

Fall 2007

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical & Computer Engineering

Program/Concentration

Electrical Engineering

Committee Director

Vijayan Asari

Committee Member

Zia-ur-Rahman

Committee Member

Jiang Li

Call Number for Print

Special Collections LD4331.E55 S516 2007

Abstract

A robust method for tracking faces of multiple people moving in a scene using a Kalman filter is proposed. This method overcomes the problem of partial and total occlusion for a short period. The method uses a combination of face detection and cloth matching to track and differentiate between people. A Template matching technique for face and a non-parametric distribution for cloth are used. Face templates are obtained from the first frame of a video sequence by applying the Viola-Jones face detection method. Cloth color distribution is obtained from people's clothes, assuming that the bodies move along with the faces. The size, top-left coordinate and velocity of motion of the detected face are used as the parameters of the Kalman vector; the predicted values are used to distinguish faces in the next frame. A threshold is set to distinguish each face from the other and to compare the face with the face in the previous frame. In the case of partial occlusion, where partial face details are lost, faces are tracked based on the Bhattacharya distance between the discrete distributions of the cloth model extracted during each frame. In the case of total occlusion where all details are lost, the algorithm uses the prediction values generated by the Kalman prediction algorithm.

The proposed method has been tested under a range of lightning conditions, change of pose, and large displacements. The results indicate that it is largely invariant to lighting changes and works well in the case of partial and total occlusion for a short period. Since the tracking is based on color, the process is computationally simple. Updating the template at discrete intervals, when the predicted values are observed to be incapable of tracking, makes the algorithm robust enough to handle sudden pose variations. Using multiple features for occlusion recovery and velocity update in the Kalman vector makes the algorithm robust with regards to occlusion and larger displacements and, also capable of working in a real time environment.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/n1ma-9q91

Download

Share

COinS