Head Pose Estimation
Head Pose Estimation
Dileep Reddy K
August 19, 2013
Head Pose Estimation
Introduction I
I Head pose estimation is the process of inferring theorientation of a human head from digital imagery
I In the context of computer vision, head pose estimation ismost commonly interpreted as the ability to infer theorientation of a persons head relative to the view of a camera
I The human head has three DOF in pose, which can becharacterized by pitch, roll, and yaw angles as pictured in Fig
I An ideal head pose estimator must demonstrate invariance toa variety of image-changing factors such as facial expression,lighting conditions and the presence of accessories like glassesand hats
Head Pose Estimation
Introduction II
Figure: Head image
Head Pose Estimation
Motivation I
I Physiological investigations have demonstrated that a personsprediction of gaze comes from a combination of both headpose and eye direction
I Consequently, to computationally estimate human gaze in anyconfiguration, an eye tracker should be supplemented with ahead pose estimation system
I People use the orientation of their heads to convey richinterpersonal information
I Like speech recognition, head pose estimation will likelybecome an off-the-shelf tool to bridge the gap betweenhumans and computers
I Has scope in HCI,Smart rooms & automative safety
Head Pose Estimation
Motivation II
Figure: Fig. 2. Wollaston illusion: Although the eyes are the same inboth images, the perceived gaze direction is dictated by the orientation ofthe head
Head Pose Estimation
Existing techniques for head pose estimation I
I Appearance template methods
I Detector array methods
I Nonlinear regression methods
I Manifold embedding methods
I Flexible models(EGM,AAM)
I Geometric methods
I Tracking methods
I Hybrid methods
Head Pose Estimation
Results
Figure: Results
Head Pose Estimation
Geometric methods for head pose estimation I
A. Gee and R. Cipolla, Determining the Gaze of Faces inImages, Image and Vision Computing, vol. 12, no. 10, pp.639-647, 1994.
Jian-Gang Wang and Eric Sung, ”EM enhancement of 3Dhead pose estimated by point at infinity”,Image and VisionComputing, Volume 25 Issue 12, December, 2007
Head Pose Estimation
Gee and Cipolla I
Assumptions
I Model ratios Rm = Lm/Lf ,Rn = Ln/Lf
I Weak perspective imaging process,frequently assumed invision research
I Facial features (eye ,mouth corners and nose point)are visibleand are given
Method
I ~n = [sinσ cos τ, sinσ sin τ,− cosσ]
I slant angle σ can be calculated using model ratio Rn and θ
Head Pose Estimation
Gee and Cipolla II
Figure: Facial model
Head Pose Estimation
Figure: method-1
Head Pose Estimation
Results of Gee and Cipolla
Figure: Results of Gee and Cipolla
Head Pose Estimation
Jian-Gang Wang and Eric Sung I
Assumptions
I weak perspective projection
I Two outer eye corners, two inner eye corners and two mouthcorners
I model ratio r = De/Dm
Head Pose Estimation
Jian-Gang Wang and Eric Sung II
Figure: Model face
Head Pose Estimation
Results of Jian-Gang Wang and Eric Sung
Figure: Results
Head Pose Estimation
Facial feature point extraction methods I
Motivation
I Face normalization is required to support face recognition bynormalizing a face for position so that the error due to facealignment is minimized
I Experiments performed by Turk and Pentland report ap-proximately 96% correct classification over lighting variation,while performance drops dramatically with orientation (85%)and size(64%) changes. For this rea- son face normalization isneeded.
I Berg proposed a rectification procedure using facial featurepoints
Head Pose Estimation
Classification
I Edge detection based methods
I Intensity image based techniques
I Hybrid methods
Head Pose Estimation
Exemplars
I Hough transform
I Deformable templates
I Active contours
I Integral projection