Research Of The Real-time Face Recognition Method Based On Embedded System

Information security and identity authentication has been an important topic in human life. With the rapid development of computer vision and machine learning, obtaining information from the image automatically and recognizing the faces became a mainstream research in the field of identity authentication. At the same time, with the advantages of small size, low cost, easy deployment and distributed computing, more and more industrial and consumer information security systems are built on the embedded platform. However, the current mainstream real time face recognition algorithms are too complex to be applied in the platforms with the low computing ability(Embedded system). Therefore, this paper launched the research from following several aspects.A multi view face classifier training model is proposed. Using the proposed model, the weak classifiers are trained into two dimensional cascade multi view face classifier. Then the final face angle is obtained by the multi view error model. This classifier can effectively detect the face images with different angles from the original image. Then, using the ellipse skin model to verify the images and remove the non-face images. Finally, using the method of face normalization proposed in this paper to relocate the face, remove the background and the minor parts of the face.Comparing with the performance of CamShift algorithm and the traditional particle filter algorithm; In view of the shortcomings of traditional particle filter algorithm, a real time particle filter tracking method which can adjust the tracking window intelligently is proposed in this paper. By the mean shift method, the sampling particles are the local optimization, which greatly reduces the number of required samples and improves the efficiency of the algorithm. At the same time, the tracking window can be adjusted automatically with no extra computation by the method. Aiming at the problem that the face cannot be updated in time by the traditional tracking method, the face tracking method based on detection and matching is proposed in this paper. In this way, the results of face detection are used as the sampling particles of face tracking to track and match the face. The method can update the face in time. And it is more accurate and less computational.The method of the real-time face recognition based on multi frame aggregation is proposed in this paper. Firstly, quick face recognition is running for the faces which are detected and tracked. And the recognition sequences are formed in accordance with the tracking correlation. Then, the final result is obtained according to the threshold judgment model. This method can effectively use the correlation between the frames in the video image, which greatly improves the accuracy and robustness of the recognition. And it meets the requirements of real time.Based on the above results, the embedded real-time face recognition system is realized in this paper. It makes full use of the advantage of multi-core processors to compute parallel. The test results show that, in the condition of moderate stream density, the system meets the requirement of real time, and has good recognition accuracy.