Model-based 3d gait biometric using quadruple fusion classifier

Abstract

The area of gait biometrics has received significant interest in the last few years, largely due to the unique suitability and reliability of gait pattern as a human recognition technique. The advantage of gait over other biometrics is that it can perform non-intrusive data acquisition and can be captured from a distance. Current gait analysis approach can be divided into model-free and model-based approach. The gait data extracted for identification process may be influenced by ambient noise conditions, occlusion, changes in backgrounds and illumination when model-free 2D silhouette data is considered. In addition, the performance in gait biometric recognition is often affected by covariate factors such as walking condition and footwear. These are often related to low performance of personal verification and identification. While body biometrics constituted of both static and dynamics features of gait motion, they can complement one another when used jointly to maximise recognition performance. Therefore, this research proposes a model-based technique that can overcome the above limitations. The proposed technique covers the process of extracting a set of 3D static and dynamic gait features from the 3D skeleton data in different covariate factors such as different footwear and walking condition. A skeleton model from forty subjects was acquired using Kinect which was able to provide human skeleton and 3D joints and the features were extracted and categorized into static and dynamic data. Normalization process was performed to scale down the features into a specific range of structure, followed by feature selection process to select the most significant features to be used in classification. The features were classified separately using five classification algorithms for static and dynamic features. A new fusion framework is proposed based on score level fusion called Quadruple Fusion Framework (QFF) in order to combine the static and dynamic features obtained from the classification model. The experimental result of static and dynamic fusion achieved the accuracy of 99.5% for footwear covariates and 97% for walking condition covariates. The result of the experimental validation demonstrated the viability of gait as biometrics in human recognition.