Fusion of face and gait for human recognition
MetadataShow full item record
A system that integrates the face, a physical biometric, with gait, a behavioral biometric, for automatically recognizing human beings effectively under a wider range of conditions than a classifier which exclusively employs only one of these biometrics, is proposed. A decision-level fusion approach is adopted where the top matches of the face classifier are passed on to the gait classifier which then determines the identity of the unknown person. For face recognition, a principle components analysis-based approach, as well as a Bayesian inference-based classifier is employed, while for gait recognition, a model-based strategy is implemented, which utilizes various gait features identified as being the most pertinent for recognition based on data collected using an optoelectronic motion capture system. The integrated system is found to outperform the individual face and gait classifiers that it is composed of, thus, demonstrating the potential of using the gait to supplement the face in scenarios where the face classifier alone does not perform well due to the non-availability of high resolution face data. During the course of this research, automated face recognition was also studied in detail, with a concentration on approaches that employ statistical dimensionality reduction techniques for this task. Experiments were conducted on some of the most widely-used methods in this category to test the recognition accuracy of these methods, and various combinations thereof, for different database sizes, images resolutions and number of bits per pixel. It was found that certain combinations of some of these techniques perform better than the individual methods that those combinations are comprised of when higher resolution face images are utilized. Furthermore, a system which utilizes different resolution versions of the whole face, as well as of various facial components, in a hierarchical manner was implemented and was found to achieve higher accuracy than its single-level counterpart which uses only the highest resolution images.