A Video Face Detection Method Using the Graph Cut Algorithm
| Autor: | Jiun-Lin Guo, 郭俊麟 |
|---|---|
| Rok vydání: | 2012 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 101 We propose a face detection system which is used in classrooms with various environments. The targets are several students in class, whose behaviors are to be observed and recorded in a classroom observation system. The feature chosen for detection is “color”, and the kernel method is the well-known graph cut algorithm. Color feature is robust against head pose changing because the area of skin regions changes little during head rotating or tilting, while other features like eyes, noses and mouth are unstable under these conditions. This character is useful for observing students’ behaviors in class since it’s usually meaningful when a student change his head pose. For example, one may doze off if he is tired, nod his head to show his agreement, or turn his head out when he distracted. These behaviors are also important events which educational researchers concern. As a result, if we can perform face detection under various head poses, then the results can be used to detect such behaviors mentioned above for further researches. To detect faces with color feature, we must choose a proper color space first, and determine the range of “skin color”. However, this kind of methods have two problems. First, the range changes with different lighting conditions and human races. Second, there are many non-human object with skin like color which affects on the precision a lot. For the first problem, we propose a dynamic learning scheme to change the skin color range frame by frame. And to solve the second problem, we propose a robust background subtraction method to eliminate non-human object. This method combining pixel based background modeling and the graph cut algorithm extracts complete foreground region from the input frame and thus avoid the effect of skin like background pixels. On the other side, since the hue values of skin color pixels are distributed concentratedly in hue color band, we apply graph cut to improve the result of skin color detection, and then collect the skin pixels for learning new skin color range. In the experiments, we set up a single camera, and there are 4~6 students in the image. Assuming an empty classroom in the beginning, the system constructs the background model first. Then, when objects appear in the image, the system will start to perform face detection. According to the experimental result, the technique proposed is robust under various head poses, and retain high precision in the low resolution images. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|