Head Pose Estimation on Top of Haar-Like Face Detection: A Study Using the Kinect Sensor

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biometrics

Background:

• Head pose estimation is vital for numerous computer vision applications, including facial expression and gesture recognition.
• Existing methods often struggle with accuracy and efficiency, particularly in handling profile views and background disturbances.

Purpose of the Study:

• To propose a frame-based head pose estimation approach utilizing the Viola and Jones (VJ) Haar-like face detector.
• To compare the effectiveness of appearance-based and depth-based features for head pose estimation.
• To introduce a novel depth-based feature descriptor for improved accuracy and reduced computation time.

Main Methods:

• Integration of the Viola and Jones (VJ) Haar-like face detector.
• Employment of various appearance-based and depth-based feature types, including Histogram of Oriented Gradients (HOG).
• Development and evaluation of a new depth-based feature descriptor.

Main Results:

• Depth data significantly improves head pose estimation accuracy compared to appearance-based features alone.
• The novel depth-based feature descriptor achieves competitive results with lower computational cost.
• Concatenation of features yielded average errors of 5.1°, 4.6°, and 4.2° for pitch, yaw, and roll, respectively.

Conclusions:

• Depth-based features, particularly the novel descriptor and HOG, are highly effective for head pose estimation.
• The proposed approach enhances robustness, correctly identifying profile views missed by frontal models.
• This method offers a state-of-the-art solution for accurate and efficient head pose estimation.