Research On Gait-based Human Identification

Biometric is a physiological or behavioural characteristic, which can be used to identify and verify the identity of an individual. As a new biometrics, gait is appealing because of it can be observed at a distance, while it is not forced special operation, difficult to conceal, and difficult to imitate. Gait recognition research currently has focused on analyzing video sequences of human walks directly. The procedure of gait recognition includes subject segmentation (extract foreground subjects from video sequences), feature extraction (extract relevant gait features from segmented silhouette sequences), and classification (classify subjects based on extracted gait features). This paper focuses on gait recognition in complex background at a distance. The main researches include gait period estimation, abnormal silhouettes restoration, especially the methods of gait feature extraction, gait feature expression, and gait feature classification.In detail, the main contributions of this paper are as follows:1. A new algorithm is proposed to estimate gait period. Swing distance of each silhouette is calculated to denote the swing extent of human body, and the periodicity of swing distances is utilized to estimate gait period. The proposed algorithm shows good adaptability to low quality silhouette images, and reduce some period error induced by other algorithms.2. A new algorithm is proposed to detect and restore abnormal silhouettes. Abnormal silhouettes are detected by silhouette distance, which is the distance of each silhouette to averaged silhouette, it emphasizes the abnormity of silhouette while neglects gait period. A detected abnormal silhouette is reconstructed by averaged adjacent image and averaged silhouette. The algorithm detects and restores abnormal silhouettes automatically without manual work. The restored silhouettes achieve much better quality and keep the consistency of gait sequence.3. In order to improve the adaptability of current algorithm to different gait conditions, this paper proposes a new gait recognition algorithm based on enhanced gait energy image. As the feature to represent gait subjects, enhanced gait energy image is constructed from two series of image templates according to the deformation of gait shape in different gait conditions. Experimental results show that this algorithm achieves a higher recognition rate than the other typical algorithms, and it show much better adaptability to the variation of surface type.4. In order to utilize more gait dynamic features, this paper proposes a new gait recognition algorithm based on primary motion contours. Three segments are extracted from silhouette contours to represent primary motions of gait. Three feature matrices are constructed based on the horizontal distances from the segmented curves to the silhouette centroid. Principal components analysis and multiple discriminant analysis are utilized to reduce redundant data and separate different classes respectively. Experimental results show that, compared with three typical algorithms, this algorithm has higher mean recognition rate and achieves better performance to the variation of carrying condition, clothing and time.5. In order to combine the static and dynamic gait features better, this paper proposes a series of gait recognition algorithms based on mean and deviation gait information. Several different basic image templetes are constructed and utilized to construct the gait deviation image which composes of the deviation of each silhouette. Gait energy image or enhanced gait energy image are utilized as gait mean image. Different fusion methods are utilized to represent gait features based on gait deviation image and gait mean image. Experimental results show these algorithms have higher mean recognition rate than other typical algorithms, and they achieve better performance to variation of other conditions without performation reduction to view angle variation. The rectified moment enegry deviation image based algorithm shows most outstanding performance, it achieves much higer mean recognition rate than other algorithms.