Three-dimensional Gait Recognition

Biometrics recognition is a method that can identify a person using physical feature or activity feature through automation technology. Gait is a moving manner of person's lower limb. As one of biometrics, gait has aroused more and more interest.Compared with other biometrics, gait has unique advantages. The advantages of gait are: it is noninvasive; it can be recognized at a long distance; it is not easy to hide; it is unique. From the standpoint of biometrics, everyone has a unique gait because its pattern is decided by individual's weight, trunk's length, customary posture and so on. This paper researches the whole gait system. We propose some new algorithms in 3D gait model's establishment and 3D gait recognition, and obtain outstanding achievements. Moreover, we have done a lot of work on the gait detection and 2D gait recognition.The establishment of 3D gait model: In this paper, we proposed a modeling method based on computer graphics and robotics. For basic geometric models, we change them by rotating, scaling, translating, connecting, and coordinate system transformation, and then the 3D gait model can be gotten. Experiment result shows, we can get arbitrary visual's gait of the 3D model through adjusting its parameters. Moreover, 3D gait model can be gotten from modeling software.The recognition of 3D gait: For the 3D gait recognition that is independent of the visual angle, six search methods including exhaustive search, multi-step search, multi-resolution search, key-frame search, evolutionary search, multi-angle search are proposed. Through synthetic comparison of different indicators of all search methods, we can find the GA is the best way. Using some searching methods to make the 3D gait model match the silhouette optimally, we get the static feature and dynamistic feature of corresponding image as the testing sample's feature. And we can get the training sample's feature by measurement. At last, sequence match method will be used to gait recognition. And the experiments made on CMU MOBO database have achieved up to 90% correction identification rate. For the 3D gait recognition relating to visual angle, first, we can get all static feature and dynamistic feature of the 3D model from 3D software as the test sample's feature. Then a network model should be established using GRNN or SVR based on the information of training samples. It presents the non-linear relation between 2D image and 3D model, so the network model can forecast the feature of testing samples. At last, SVM will be used to gait recognition. In experiment, we use the model figures of POSER and the GRNN network, and get a 100% correction identification rate by making use of the methods mentioned above.The detection of gait: We propose two approaches that called edge-frame difference and frame difference-background subtraction, and do real-time gait detection in our laboratory using CCD camera. Experiment proved these methods that mentioned above can satisfy the need of real-time processing, and the detect results is also relatively good.The recognition of 2D gait: In this paper, we propose many methods of 2D gait recognition such as gait recognition based on chain code, gait recognition based on wavelet descriptors (WDS), multi-resolution local moment feature for gait recognition. By utilizing these proposed approaches, and testing on the CMU MOBO database, we have achieved comparatively high correction identification rate.Using Matlab and Visual C++, we combine these new methods mentioned above with conventional methods which is used to image processing, and use them to 3D gait recognition. 3D gait is a combination of image processing, video processing, computer vision, pattern recognition, machine learning and other information technologies, and it has a broad prospect.