Human Identification Based on Gait and Soft Biometrics

Abstract

Nowadays, one of the most important and challenging tasks in Biometrics and Computer Vision is the automatic human identification. This problem has been approached in many works over the last decades, that resulted in state-of-art methods based on biometric features such as fingerprint, iris and face. Despite the great development in this area, there are still many challenges to overcome, and this present work aims to present an approach to one of them, which is the automatic person identification in low-resolution videos captured in unconstrained scenarios, at a distance, in a covert and non-invasive way, with little or none subject cooperation. In scenarios like this, the use of classical methods may not perform properly and using features such as gait, can be the only feasible option. Gait can be defined as the act of walking. Early studies showed that humans are able to identify individuals by the way they walk, and this premise is the basis of most recent works on gait recognition. However, even state-of-art methods, still do not present the required robustness to work on a productive environment. The goal of this work is to propose an improvement to state-of-art gait recognition methods based on 2D poses, by merging them using multi-biometrics techniques. The original methods use gait information extracted from 2D poses estimated over video sequences, to identify the individuals. In order to assess the proposed extensions, two public gait datasets were used, CASIA Gait Dataset-A and CASIA Gait Dataset-B. Both datasets have videos of a number of people walking in different directions and conditions. In the original and in the extended method, the classification was carried out by a 1-NN classifier using the chi-square distance function.