Research On Motion Human Face Detection In Digital Videos Surveillance System

Motion human face detection has a wide use in digital video surveillance system and has also drawn much attention from fields such as AI and PR. It is obvious that the research on face detection in videos is very complicated, due to its involvement in two kinds of expertise: segmentation of moving objects and human face detection. So far the general achievement is still unsatisfactory.This paper presents a new motion face detection system based on the structure and theory of digital video surveillance system. Combined with the present research level at home and abroad, the proposed system is built upon the detection of the moving objects and hierarchical neural network model, which is composed of three fundamental parts: segmentation of moving objects sub-system, candidate face preprocessing sub-system and accurate face detection sub-system.The main works and conclusions as follows:(1) In the segmentation of moving objects sub-system, an improved optical flow algorithm for segmenting the available moving object based on the analysis of traditional segmentation algorithm is proposed, which is very robust with strong noise-resistant abilities. Furthermore, the part of this algorithm for smoothing and segmenting the candidate face area makes best use of Closing in the language of morphological processing along with the filtering in image processing to eliminate available noises brought about by the above motion segmentation algorithm and finally merge the scattered small areas into one. (2) In the candidate face preprocessing sub-system, the algorithm proposed in this paper first filters the candidate face so as to eliminate the noises, and then extends in gray level up to adjust the contrast of image, and finally uses histogram equalization to adjust the mean value and square difference of images, realizing the gray level normalization.(3) In the accurate face detection sub-system, a new two-stage neural network model is proposed. The system utilizes the gray level and component of the original color image as the features of neural network classifiers and cascades two gray level-based and chrominance component-based neural networks respectively. Experimental results show that this method can detect human face effectively and fast.(4) The "segmentation-searching" mode in this paper solved well the problems of vast computation and large searching space always involved in the traditional detection algorithms. It is also a practical method to detect the motion human face in digital videos combined with the motion information, gray level information and skin color information.