Gait Recognition Based On Feature Template Construction Using Linear Interpolation

With the increasing demand of intelligent surveillance in safe and sensitive environments, human identification often needs to be carried out at a long distance. The expectation is that the authentication technology gives adequate forewarning thus preventing the probability of risk occurring. Gait recognition aims at identifying a person by the way he or she walks. Among other biometric features used for person identification, gait has prominent advantages of being non-contact, non-invasive, unobvious, low resolution requirement, insensitive to environment and it is a unique biometric feature for human identification at a distance. Therefore, it has received researchers' great interest.This dissertation concentrates on, human body detection and segmentation, feature extraction, and recognition using digital video image samples. The research focuses as follows:The existing gait databases founded by well-known research institutions are summarized briefly. The HEU gait database was established by our research team. It came from four individuals each walking in eight different directions. The camera was mounted and their video sequences were collected at a dipping angle.To achieve proper human body detection and segmentation, three methods namely: optical flow, frame difference and background subtraction were explored. After analyzing the effectiveness and complexity of these three methods, direct background subtraction was used for the indoor scenario, while background subtraction with an updating background was used for the outdoor environment. Then morphology transformation was used to fill cavities in a single binary image. Connected components analysis was performed to obtain the silhouette of human body; also the standardized and centralized image was realized by geometry conversion of graphs in order to eliminate the effect of different image scales on recognition performance.Feature extraction is the key issue in this research, and is the most significant factor. For each gait sequence, the transformed gait images are used to construct templates including structural and dynamic gait features. The real continual gait sequence is obtained using linear interpolation of discrete gait sequence images; thus forming a single vector or matrix with attributes of gait characteristic. This avoids the dynamic time wrapping complex process during recognition. Classical PCA, 2DPCA, C2DPCA as well as GLRAM dimensionality reduction methods were analyzed extensively. Three simple and effective algorithms which unify the actual characteristic forms were proposed namely; linear interpolation of sole angle projection and linear interpolation of combined angles projection based on integration with PCA respectively, and linear interpolation of FanBeam projection based on integration with GLRAM. In order to confirm the effectiveness of features extracted, the nearest neighbor classifier was applied at the recognition stage. Although the merit of angle projection lies in low dimensional features and non-intensive computation, its shortcoming is that the recognition rate is not high. The recognition rate of sole angle projection is lower than that of combination angles projection. As shown by the experimental results, the best result was obtained when w is 5pi/16T in the feature template Q(w,t).Because of the limited features from combining 0°,45°,90°and 135°angles projection, FanBeam projection based on linear interpolation integrating with GLRAM method which has abundant features was used in the feature extraction process. After balancing FanBeam projection time consumed, its result and reconstruction result, the parameters were determined. Then, GLRAM was employed for further dimensionality reduction. Finally this method was validated by experiments and achieved the best recognition rate of 0.8871 in the CASIA(B) gait database and 100% in the HEU gait database. The parameters are w=5pi/16T ,D=49 and s=2.