A New Method for Gait Recognition Using 2D Poses

Abstract

Over the last decades, biometrics has become an important way for human identification in many areas, since it can avoid frauds and increase the security of individuals in society. Nowadays, most popular biometric systems are based on fingerprint and face features. Despite the great development observed in Biometrics, an important challenge lasts, which is the automatic people identification in low-resolution videos captured in unconstrained scenarios, at a distance, in a covert and noninvasive way, with little or none subject cooperation. In these cases, gait biometrics can be the only choice. The goal of this work is to propose a new method for gait recognition using information extracted from 2D poses estimated over video sequences. For 2D pose estimation, our method uses OpenPose, an open-source robust pose estimator, capable of real-time multi-person detection and pose estimation with high accuracy and a good computational performance. In order to assess the new proposed method, we used two public gait datasets, 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 our new method, the classification is carried out by a 1-NN classifier. The best results were obtained by using the chi-square distance function, which obtained 95.00% of rank-1 recognition rate on CASIA Gait Dataset-A and 94.22% of rank-1 recognition rate on CASIA Gait Dataset-B, which are comparable to state-of-the-art results.