Surveillance Video Based Face Recognition And Moving Object Segmentation

Researches of face recognition and moving object segmentation, as important sub-fields of computer vision and pattern recognition, have been of great academic and practical significance. The use of spatial and temporal information in video has effectively overcome the difficulties existing in the video-based face recognition, such as low resolution of face images in video, large variations of face scale, radical changes of illumination and pose, as well as occasionally occlusion of different parts of faces. In addition, the result of segmentation is the foundation of a range of high-level scene and object analysis. On the basis of the existing work in the related field of this study, the present paper proposes a new video face recognition algorithm and moving object segmentation algorithm, and analyses several typical face recognition approaches with good illumination robustness in detail.Video-based face recognition has become popular recently. The paper proposes a novel method to reconstruct a 3D model of a human face from video sequences to perform video-based face recognition. On the training stage, the final 3D face model is obtained by pose estimation based on feature points location algorithm of a modified active shape model (ASM), together with a generic model of human face and the texture mapping using the front facial image. On the recognition stage, by estimating the pose of detected facial images from the input videos and adjusting the pose of 3D face model stored in the database, then the highest score among the likelihood scores provided by 2D empirical mode decomposition (2D-EMD) and PCA dimensionality reduction features extraction estimates the identity of the test video sequence. Experimental results illustrate that the proposed algorithm outperforms the selected benchmarks representative of existing techniques.The unavoidable illumination variations have direct negative effects on the face recognition accuracy in the surveillance videos. And the differences among face images of the same person in different lighting conditions are much more obvious than those among face images of different person under the same conditions. Therefore, face recognition research with good illumination robustness has been an important research topic in the field of pattern recognition. In the light of the existing robust illumination approaches, there is a comprehensive and in-depth study of face recognition algorithm based on sparse representation, phase congruency feature and local binary patterns feature in this paper. Through the experiments carried out on the ORL, AR and Extended Yale B face databases, and also based on the illumination robust feature extraction and utilizing principal component analysis (PCA) to reduce the feature dimension, the final face recognition and classification via distance matching shows that these three kinds of face features can effectively overcome the factor of illumination.Video-object segmentation is an important research area of computer vision with applications of video surveillance, traffic monitoring, multimedia, etc. Most of the recent methods work on image pixels or color segments which are with high time cost and computational complexity. In this paper, we proposed a novel video-object segmentation approach inside surveillance videos, which combine optical flow motion estimation method with noise-eliminating and the graph-based image over-segmentation approach. During the process, the motion information of moving object is detected by optical flow method among successive frames. And by using digital filter to remove the noise presented in the optical flow data, the positions and boundaries of moving object can be better segmented. Then, the object over-segmentation is achieved by the graph-based segmentation method. Finally, using the graph structure integration framework and referring to the positions and boundaries of moving object in the video sequences, the over-segments of moving object can be merged to produce a proper segmentation result. With a number of test video sequences, the experiment results show that the proposed approach can achieve a good performance in video-object segmentation.