A Model-Based Particle Filter for 3D Head Pose Estimation
 
 
 
Nian-Tzu Gau(高念慈)
中華民國 100 年 6 月
輔仁大學---智慧型系統實驗室---雲端視覺組
 
 
 
 
Abstract
    ‧Head pose estimation is a technique that determinate the orientation of face. The orientation of human face is a important information, face is a significant symbol that show human attention and behavior. For estimating the pose of head, tracking the feature points on face is very important. Particle filter is a tracking algorithm that alternative of extend Kalman filter, it has been widely used for solving tracking problem. It predict a moving object location from observation value that contains noises. In this paper, we propose a model-based particle filter that tracks the feature point on the face and fits by AAM. the proposed model-base particle filter that use non-linear regression analysis to train a state transition model to make the state transition more efficiently. The experimental result show that model-based particle filter have better head pose estimation than classic particle filter.
 
Method
    ‧ The system framework is divided into two part: train and testing. In training part, the facial landmarks are trained by the non-linear regression analysis, we can acquire the parabolic parameter after training. In testing part, we proposed a model-based particle filter to track the facial feature point, and the model parameter is acquired from the result of non-linear regression analysis. After AAM fitting process, the facial feature points and POSIT algorithm are used for head pose estimation.
    Our proposed state transition training framework, we extract the facial landmark from the training images that contain night direction of face pose. And we divide landmark into four parts: chin, eye, eyebrows and mouth. The four parts are trained individually by using regression. After the training is done, we obtain the coefficient of curve.
    In comparison of state transition on particle filter and model-based particle filter. the samples are generated and diffused independently from t to t+1. But the model-based particle filter generates samples dependently. For constraining the samples, we use regression analysis to construct the relation of samples.
 
 

 
 

 

 

 

 

Conclusions
    ‧This paper proposes a model-based particle filter that tracks feature points on face, and head pose is estimated by POSIT algorithm. The proposed model-based particle filter use the non-linear regression analysis to train state transition model that could reduce the number of particles used in the tracking process. The experimental result show that model-based particle filter have better head pose estimation than classic particle filter.
 
 
Demo Video
 
 
 
 
 
輔仁大學---智慧型系統實驗室