Research On Biometric-based Identity Recognition In Videos

Identification technology is an important research topic in the field of pattern recognition, and because biometric identification can recognize the person only based on the body itself, which can avoid the problems of the traditional identification methods, such as crack and steal, and has high accuracy and safety. The biometric-based identity recognition in videos is the research of face recognition and gait recognition, which can be collected more friendly and applied more widely than other biological characteristics. It has become a research hotspot in recent years. This paper focuses on the research of biometric identification in video sequences, and the main contributions are summarized as follows:Firstly, to address the mismatch of the gallery and the query in the still-to-video face recognition, an improved still-to-video face recognition algorithm based on sparse representation is proposed. Faces in the video sequence have the alignment difficulty because of the detection results or the expression change. To enable the face alignment, it uses gradient variance to locate the main contours of the face. Under the static condition of the camera, the motion of the human face leads to the blur, resulting in the obvious decrease of recognition rate. In order to robustly recognize face in motion blur condition, a multi-scale rectangle filter is used to create a dictionary, and a multi-scale blur feature is added, which makes the dictionary anti-blur. Since the video sequence contains a lot of redundant information, the key frame extraction method based on cross correlation coefficient clustering is proposed in this paper to deal with the large amount of computation.This method can adaptively determine the number of clusters without pre-set, and it can represent all the information in the video sequence, eliminating the redundant information and reducing the recognition time. The results show that the proposed algorithm has significant improvements than the neural network and the support vector machine.Secondly, because of the changes like illumination and expression of the face images in video sequences, the data distribution of faces is dispersed and the between-class scatter is less than the within-class scatter. Also the traditional classifier is more sensitive to data,and is bad at generalization ability. In order to make full use of the multi-frame images and motion information in video sequences, considering the scatter information and the generalization ability of classifier, a novel video face recognition algorithm combining scatter constraint and hierarchical cascaded classifier is proposed. First, it divides themulti-frame images in a video sequence into several clusters based on hierarchical cluster.Then, the rest samples of the gallery are identified by the minimum distance method. If the recognition result is wrong, the next layer classifier is established for the misclassified samples and similar samples. The final classifier is cascaded through constructing hierarchical classifiers. Finally, it recognizes test video sequence through created ensemble hierarchical classifier. This method creates the classifier layer by layer to subdivide the data distribution, increasing the ratio of between-class scatter and within-class scatter, and focuses on the easily misclassified samples learning, which improves the generalization ability of the classifier. Experimental results show that the proposed algorithm can effectively deal with the problem of face recognition in video.Finally, in order to improve the recognition rate and make full use of gait information,a gait recognition algorithm combining discrete cosine transform and linear discriminant analysis is proposed in this paper. First, extract the frequency domain information from the discrete cosine transform. Then the data is mapped by linear discriminant analysis, and finally achieve identification.This method can effectively distinguish the high and low frequency information from the gait features, and enhance the discrimination of features.Experiments are conducted in the database of Chinese Academy of Sciences Institute of Automation. The results show that the proposed feature extraction method is superior to the others and can achieve higher recognition rate.