Research On Real-Time Face Detection Algorithm Based On Video

Face detection is the first step of the face recognition systems, the purpose of which is to locate and extract the human face from the background. It is a active research topic in computer vision and pattern recognition and is used in people’s daily life widely, such as video monitoring, human computer interaction, image retrieval, video conference, authentication, virtual reality and so on. So it has broad application prospect and social benefit. With the promotion of applications, all kinds of new methods and technologies are introduced into face detection to adapt to more complex environment. Due to various changes of the imaging background, such as weather, lightness, shadow and background interference effects, until now there is not an method adapted to all situations. So research of a robust and highly accurate face detection algorithm with high performance is a great challenge in this research area.On the basis of reviewing the previous works and doing research on a variety of detection and tracking algorithms and their application occasions, the thesis proposes improvement on some detection and tracking algorithms, which combines face detection with some tracking algorithms of good performance and solves the problem caused by partial occlusion, facial expression, posture and illumination and so on. The proposed methods can meet the real-time requirement of video surveillance systems and realize face detection, location and tracking in the videos. The main work is described as bellows:(1) To overcome the defect of AdaBoost algorithm in time consuming caused by extracting a large number of Haar features in the training phase, this thesis uses the Circular symmetry Gabor transform (CSGT) features to replace the Haar features, extracts the local maxima of the transformed image as features and constrains their uniform distribution to train the classifier. The CSGT inherits the good properties of Gabor wavelet. At the same time, it has strong performance and strict rotation invariance. So the improved AdaBoost algorithm has better robustness to the change of expression and position. Finally the experiment results indicate that this method obtains better testing effect. It is also found that, on the basis of improving the training speed, it sacrifices some detecting time. However, following the scientific development, such as the application of the multi-core processors and accelerated hardware, the detection speed can be greatly improved.(2) After analyzing the face detection based on skin color segmentation, we combine it with AdaBoost algorithm to face detection. According to complex background images, we use the skin color segmentation to extract the candidate skin color areas, and then use the face model proportion to remove some non-face areas, and finally, use the trained AdaBoost classifier to detect faces. The test results show that, compared with AdaBoost algorithm, this method improves not only the detection precision, but also the detection speed.(3) Aiming at the object detection in moving videos, the thesis introduces motion estimation to extract the moving areas, and detects faces using the above face detection algorithm. This method can eliminate the disturbance of background motion and has encouraging detection speed and real time effect.(4) Besides detection algorithm, the thesis introduces the SIFT (Scale-Invariant Feature Transform) method using object matching up to tracking result, which extracts features with invariance on scales, rotation and translation, and has better robustness to illumination variation, affine transform and3D projection transform. In this thesis, the method is slightly modified. Because of the high resolution of images, it extracts too many feature points with the large amount of calculation. We add distance constraint for feature points after removal of edge points and instable points, to get uniformly distributed feature points and decrease the matching time with the same match probability.