Video surveillance face recognition by more virtual training samples based on 3D modeling

Abstract

Video surveillance has been applied in more and more fields for security in last decade years, video-based face recognition therefore became an important task of an intelligent monitoring system. However, among these captured video faces there are many non-frontal faces. As a result, the state-of-art face algorithms would become worse when they were employed to recognize video faces. On the other hand, it was a common phenomenon especially at video monitoring field that only one training sample per person is gained from their identification card. The single sample per person (SSPP) results in effecting even not taking advantage of some fine algorithms such LDA. In order to effectively improve the correct recognition rate of multi-pose face recognition with a single frontal training sample, this paper proposed a face recognition algorithm based on 3D modeling. In the proposed algorithm, firstly a 2D frontal face with high-resolution was taken to build a 3D face model, and then several virtual faces with different poses were produced from the 3D face model. At last, both the original frontal face image and virtual face images were put into a gallery set. The algorithm was evaluated on SCface database using traditional PCA and LDA methods. The result showed that the proposed approach could effectively improve video face recognition rate and the correct recognition rate went up about 13% by LDA compared with traditional PCA. Therefore, the method that was proposed to create virtual looking down training samples was an effective algorithm and could be considered to apply in intelligent video monitoring system.