Tracking and Re-identification of People Using Soft-Biometrics

The goal of this work is proposing a method of biometric identification using soft-biometrics, that aims the extraction of physical characteristics and estimation of the pose as unique traits of each individual, to name and trace that specific person trough the scene. In this work we partially used the public database CASIA Gait Database-A, which has several frames of people, already classified, walking in different directions and angulations, along with a set of silhouettes that were extracted from these scenes and the background used at recordings. Besides, we used a private database of the project sponsor, Petrobras, containing videos of security cameras used to demonstrate the daily routine of workers at an oil platform. The biggest challenges of performing biometrics in this dataset are the quality of the provided images and the heavy clothing used by the workers on the platform, that often hinders the processing quality of the algorithm, explaining why we chose to work with soft-biometric. The algorithm used in this method is PifPaf, made to estimate the human pose and extract features and capable of performing the detection in environments with noises, low illumination or low resolution. With its help, we mean to extract parts of the workers bodies in the private database and from the actors in the scenes from the CASIA Gait Database-A. For our methodology we used the Euclidean and city block distance calculations, obtaining 70% hits with a combination between the PifPaf algorithm and Euclidean distance.