Embedded Video Surveillance System Based On Face Recognition

Nowadays, video surveillance spreads to every corner of the society and the intelligent video analytics is gradually becoming the development orientation of video surveillance. In this paper, we has described an embedded security and attendance system based on face recognition. It has the extremely significant significance for the monitoring of the people and the protection of property.From the perspective of practical application, we are aimed to develop a hardware resource limited attendance and surveillance system to monitor the people going through the surveillance scenarios based on a surveillance development board named HI3520 D.Main work includes the following aspects.1, Overall scheme and architecture design of the embedded system. Embedded surveillance system can work on F(front) mode or S(server) mode.The software of system consists of five modules:realtime video,motion detection,face detection and recognition,user listening and interactive communication.By analyzing the interactive communication,we draw the communication processing chart and promote the protocol.2, Object detection.At first, we roughly determine where object comes by analyzing the data from API of HiSilicon chip’s VDA module.Then, through the API of HiSilicon,we get the macro block SAD(Sum of Absolute Difference) of current image and background. Finally, by scanning the SAD, we locate the object. Besides, we can estimate object’s state of motion by analyzing the locations of object.3, Face positioning and snaping.Firstly, we detect face by Haar features, if failed, then detect by LBP features. By combining two algorithms, we can reach a lower omission ratio.4, Face recognition. Firstly, we deal the faces with size normalization and histogram equalization. We descending order images by definition according to the energy gradient, and retain the first three pictures.Then, we finish the face recognition by LDA algorithm combining with three images.5, Stranger descrimination.Through a large number of face data, we aquire everyone’s face recognition Euclidean distance distribution. Each person’s Euclidean distance approximate accord with normal distribution. By sacrificing some accuracy, we realize stranger descrimination.Besise, the intelligent surveillance system can also combine the method proposed in this paper with gait, audio, etc to improve the reliability of the security and attendance system’s reliability.