The Online Learning Algorithm For Face Motion Tracking

Computer vision like the eyes of computer. More exactly, we use the web camera and computer to track、detect and analyse the object instead of our eyes. Recent years has witnessed an alarming increase in video-based Face Tracking in the field of computer research. This technology took advantage of many subjects, such as Image Processing, Pattern Recognition, and Artificial Intelligence. It has been widely applied in Human-Computer interaction and intelligent video surveillance.In face motion tracking, the target is firstly located by hand in the first frame, then, a face detection or tracking approach is employed to track the face. Face detection approach trains samples to obtain the visual prior of the target. The learned visual prior can compactly represent the variance of object appearance, which enables object detection to track object precisely under simple background. Howver, the performance will dramatically decrease under occlusion or illumination variance. Object tracking exploits the online information about object and background to enhance the tracking performance. However, track drift and track loss may appear in the long sequences of face video tracking. Since Face Tracking and Face Detection have their own advantages and disadvantages, therefore, these two approaches can help each other during the tracking. Inspired by this, the paper investigates the combination of Face Detection and Tracking after thoroughly studying the approaches of Face Detection and Tracking. In the paper, a PCA+TLD algorithm for face motion tracking is proposed, which can improve the accuracy and robustness during the tracking.This paper mainly does the following four tasks:1. Introduce some tracking algorithms under complex background, and detail three typical tracking algorithms:Optical Flow, Partile Filter and Mean Shift.2. Introduce some common features used in object detection. Focus on the principle and method of PCA for face detection.3. Introduce the TLD, focus on the On-line P-N learning.4. A novel PCA embedded TLD approach was proposed for face tracking. The proposed approach can enhance the accuracy and stability in the face tracking.