Identificação de indivíduos pela dinâmica do caminhar

Abstract

The security systems based on identification of individuals through biometric characteristics have required that the programs responsible for this function are as reliable as possible. Biometrics modalities such as face, iris and fingerprint require sophisticated resources and the straight help of the individual which wish to recognize. Based on this context, it was sought a way to minimize these requirements through the biometric characteristics of walking. The human walking is a characteristic of each person, and this aspect can be used favorably in recognition systems. In this work was developed a system for identification of individuals by the dynamics of gait, using a holistic approach to compose the vectors of characteristics. For this it was used a robust background estimator, the LMedS, to do the extraction of silhouettes. After this stage, it was done some filtering on the images in order to improve the quality of silhouettes captured, and the measure of the width of these silhouettes were used as characteristic of each class of people. Size reduction methods as the PCA, LDA and POV were tested so that the feature vectors became more representative and that the system processing time would be reduced at the later stage. As a classifier, it was used the HMM, because it is possible to relate the states of a gait cycle to the states present in HMM. In this case, as each person has a specific way of walking (for example, some walk faster, others walk more slowly), this classifier fitted very well to the research proposal. The results showed that the LDA feature extraction is the most satisfactory because giving, with the HMM classifier, the most expressive hit rates. Given the current trends in biometric fusion, gait shows to be very attractive for use in bi-modal systems or even alone.