Gait Recognition Based On Statistical Characteristics Of Image Sequences

Gait, which captures the manner of human walking, has been considered to be the most potential biometrics in the area of intelligent visual surveillance at a distance. It has the merit of being non-invasive, hard to conceal, being readily captured without physical contact, and executable from a distance. Gait recognition is the technique to identify humans by their walking pattern, and it has gained increasing interest from computer vision researchers recently. The study of gait recognition technique could promote the development of the theory of computer vision and pattern recognition. It also has extensive applications such as the video surveillance in security-sensitive places, access control for special occasions and assistance to catch criminals, etc.Video-based gait recognition technology mainly concentrates on gait silhouette images composed of moving pixels. Because of various system noises such as camera shake, variation of illumination and occlusion, the quality of the detected gait silhouette is always very low. Rather than extracting or matching features from single silhouette image, this thesis deals with the entire images during a certain time interval through statistical techniques. This study can be concluded to be statistical gait recognition, because the spatial-temporal characteristics of gesture changing are represented by certain kind of statistical measurements in this research. The main purpose of this work is to improve both the identification performance and the verification performance of gait recognition algorithms, and also to furtherance the practical use of gait recognition techniques. This thesis is supported by the National Natural Science Foundation of China under grant No.60675024.In detail, the main contributions of this thesis are as follows:1) In order to provide a better shape presentation of gait silhouettes, we propose a probabilistic model-based silhouette refining algorithm to fill in holes and recover missing parts automatically. First, all the raw silhouettes are evaluated by a quality detection algorithm and those silhouettes with low quality can be automatically detected. Then, the distorted silhouettes in a particular sequence are refined by an individual model which is trained in advance by the current sequence. If the quality of some silhouettes is still low, then those silhouettes can be further refined by a population model. The experimental results show that the refined silhouettes not only have a better presentation but also help improve the recognition performance of the existing gait recognition algorithms. The characteristic of this refining algorithm is that it only refines the silhouette in bad quality, and the between-class similarity can be enhanced because most silhouettes are undergoing a self-updating procedure.2) The spatial-temporal variations of silhouette contours capture the walking patterns of human being, and the mean shape of a gait image sequence can be obtained by applying the method of Procrustes shape analysis. A novel shape descriptor, shape context, is introduced to depict the distribution of the sample points on the mean shape. Shape context uses normalized histogram bins to describe the relative spatial relationship of the boundary points and offers us a powerful gait feature representation. The computation cost can be decreased by a fast point matching strategy. The experimental results indicate that the classification performance of the mean shape representation can be further promoted by introducing this shape descriptor.3) On the basis of mean shape representation of gait sequences, this thesis proposes a novel statistical gait recognition algorithm based on tangent angle features. The tangent angle of a certain point on the mean shape is defined as the corresponding angle of the tangential vector at that point in vector space. The tangent angle is considered to reflect the local appearance and tendency at that particular point and is treated as a local discriminative feature called tangent angle features (TAF). The local tangent angle dissimilarity is used to measure the distance between two different TAFs, and the simplest standard classifiers are used to distinguish different patterns. The experimental results reveal that, the proposed algorithm outperforms other existing approaches in terms of recognition accuracy.4) The mean silhouette of the gait sequence, which is also called gait energy image (GEI), can be decomposed into eight bit-planes. We consider that some bit-planes have more structural characteristics, while the others have more detailed features. Towards combining those bit-planes according to different weights respectively, we can get structural image and detailed image of the original GEI. They represent the structural information and detailed information respectively. A virtual gait energy image (VGEI) can be obtained by integrating the structural image and detailed image in complex space. The generalized PCA is applied to VGEI to reduce the dimensions. The classical Euclidean distance is used to measure the similarity of different gait features, and the nearest neighbor classifier is adopted to discriminate different patterns. The experiments on CASIA database testify the effectiveness of the proposed algorithm.5) From the perspective of texture analysis, we try to extract texture features from GEI to achieve a gait recognition algorithm. The extracted texture features include local range, local standard deviation and local entropy which reflect the local variability of the intensity values of pixels in GEI. The corresponding local range images, local standard deviation images and local entropy images can be trained and applied to accomplish the gait recognition task. They show exciting identification performance and verification performance on CASIA database. It has proved that the texture features have stronger discriminative power than the original GEI.6) In order to improve the identification performance and verification performance, we investigate the algorithm that utilizes the theory of information fusion. The gait features can be fused in two aspects:one is contour-based gait feature which includes the mean shape, the shape context matrix and the tangent angle features, and the other is area-based gait features which include the local range, local standard deviation and local entropy. Large amount of experiments have shown that, both the identification performance and verification performance can be promoted to a certain degree by fusing multiple features, and the fused features outperform any single feature when it is used individually.