Research Of Human Gait Recognition Via Deterministic Learning Theory

Recognition by gait is a new and attractive research field for the computer visionand biometrics recognition technology. Its aim is to recognize people or detect physio-logical, pathological and mental characteristics by their walking style. Compared withother biometrics recognition technologies, such as fingerprint, iris and face, gait is hu-man's explicit and dynamic representation which is closely related to the information ofspatial-temporal waking movement. Gait recognition is the unique perceivable biomet-rics at a distance and has its predominance among other identity recognition technologieswhich are based on static features, because it has the advantages of being noninvasive,requiring little about the quality of video and difcult to disguise. It is an attractive bio-metrics recognition technology in the intelligent monitoring system. The key point of gaitrecognition is to extract suitable static or dynamic features or their combination whenanalyzing the human walking videos captured from cameras. Gait signatures are thenderived and integrated with classifiers to achieve gait recognition. It contains many kindsof techniques, such as computer vision, pattern recognition, video and image sequencesprocessing and so on. Based on its strong points and importance both in theoreticalresearch and practical application, gait recognition will be further studied based on thedeterministic learning theory in this dissertation. The main contribution and innovationof this dissertation are summarized as follows:1. A new model-based approach for human gait recognition is presented. The ap-proach consists of two phases: a training (learning) phase and a test (recognition) phase.In the training phase, side silhouette lower limb joint angles and angular velocities areselected as gait features and extracted from the human walking image sequences by us-ing image processing techniques. A five-link biped model for human gait locomotion isemployed to demonstrate that functions containing joint angle and angular velocity statevectors characterize the gait system dynamics. Due to the quasi-periodic and symmet-rical characteristics of human gait, the gait system dynamics can be simplified to bedescribed by functions of joint angles and angular velocities of one side of human body,thus the feature dimension is efectively reduced. Locally-accurate identification of thegait system dynamics is achieved by using radial basis function (RBF) neural networks(NNs) through deterministic learning. The obtained knowledge of the approximated gaitsystem dynamics is stored in constant RBF networks. Hence, time-varying gait dynam-ical patterns can be efectively represented by the locally accurate NN approximations of system dynamics, and this representation is time-invariant. A gait signature is thenderived from the extracted gait system dynamics along the phase portrait of joint anglesversus angular velocities. A bank of estimators are constructed using constant RBF net-works to represent the training gait patterns. In the test phase, by comparing the set ofestimators with the test gait pattern, a set of recognition errors are generated, and theaverage L1norms of the errors are taken as the similarity measure between the dynamicsof the training gait patterns and the dynamics of the test gait pattern. Therefore, thetest gait pattern similar to one of the training gait patterns can be rapidly recognizedaccording to the smallest error principle.2. A new approach for human gait recognition based on dominant gait systemdynamics is presented. Human gait locomotion is simplified as a5-link biped model.With the side silhouette lower limb joint angles being selected as the state vectors, locally-accurate identification of the dominant gait system internal dynamics can be achievedwithout using the joint angular velocities. It can not only reduce the feature dimensionbut avoid the generation of observation error without using the high-gain observer toestimate the joint angular velocities. Combined with Labview software and the DawningServer, we construct a gait recognition prototype. The recognition speed under multi-pattern is promoted by using the parallel programming mode of Labview and the multi-core CPU of Dawning Server, which makes the gait recognition system be close to thereal operation as much as possible.3. A new approach for human gait recognition based on feature fusion is presented.The vertical coordinates of the center of mass of the side silhouette are selected as thestatic gait feature, side silhouette lower limb joint angles are selected as the dynamicgait feature, both of which are combined together as a new gait feature to overcomethe problem generated by using single feature. Locally-accurate identification of thegait system dynamics is achieved by using RBF NNs. The obtained knowledge of theapproximated gait system dynamics is stored in constant RBF networks. A gait signatureis then derived from the extracted gait system dynamics along the phase portrait of jointangles versus the vertical coordinates of the center of mass of the side silhouette. A bankof estimators are constructed using constant RBF networks to represent the training gaitpatterns. In the test phase, by comparing the set of estimators with the test gait pattern,a set of recognition errors are generated, and the average L1norms of the errors aretaken as the similarity measure between the dynamics of the training gait patterns andthe dynamics of the test gait pattern. Therefore, the test gait pattern similar to oneof the training gait patterns can be rapidly recognized according to the smallest error principle.To verify the efectiveness of our gait recognition algorithms, a lot of experimentshave been performed in the CASIA gait database. Then, the experimental results areanalyzed and discussed in depth. Experimental results show that the proposed algorithmshave higher recognition rate.