Research Of Face Detection Based On Embedded Smart Surveillance System

As an important feature recognization technology, the Face Detection and Recognization is always the hot point in the area of computer vision and pattern recognization. With the rapid development of Embedded System and the wide application of Intelligent Video Surveillance, it has become a new topic in Face Detection and Recognization that how to develop the technology in combination with Embedded devices. In this article an Intelligent Video Surveillance System has been developed based on the embedded platform which was constructed by the Video Codec Chip GM8180, the realization of this system brought in the Visual Selective Attention Model and meanwhile adopt the Adaboost Face Diction Algorithm and the Camshift Face Tracking Algorithm.Firstly, a research on model concept, feature classification, extraction pattern etc of Visual Attention Model has been performed on this paper, as well as analysis on the characteristics of both bottom-up model and top-down model. At the same time, a comparison has been made between different algorithms of feature selection and feature area extraction. The Saliency Map Theory has been applied in extraction of video saliency area, by which the extraction of feature area maps and generation of saliency map was completed. Then the cursory face detection was finished through threshold decision.Secondly, the accurate face detection was worked out by use of Adaboost face detection algorithm based on Haar-like features. Specific to the embedded platform on which the system work, several optimization strategies has been adopted in this paper, including transformation from float point to fixed point, parameter optimization and the search based on saliency area, etc. By this way, the speed and efficiency of detection were improved obviously on the target embedded platform, and the real time detection based on saliency area was implemented.Finally, a comparison among various popular algorithms of face tracking nowadays, according to real application requirements, the Camshift algorithm was adopted, combined with the Adaboost Face Detection Algorithm. The first step is locating the rectangular face searching box, and then, a hue probabilistic model of face is constructed to transform the input video frame into hue histogram. Through locating the mass center and adjusting the position of the search box, Camshift performs face tracking smoothly.By the end of the research, a face detecting and tracking system has been implemented successfully on GM 8180 platform, which is able to search face position accurately and track face continuously. The results show that, the introduction of Visual Attention Model allow the system detect and track the saliency face in video frame firstly, while the non-face areas are excluded. Compared to traditional face detection algorithm, the Adaboost Face Detection based on visual attention reflects distinguishing priority, which is more in line with the function concept of the intelligent video surveillance system, and the Camshift Tracking Algorithm generally meets the requirements of real time tracking system.