Person re-identification from CCTV silhouettes using generic fourier descriptor

Abstract

Person re-identification in public areas (such as airports, train stations and shopping malls) has recently received increased attention within computer vision research due, in part, to the demand for enhanced levels of security. Re-identifying subjects within non-overlapped camera networks can be considered as a challenging task. Illumination changes in different scenes, variations in camera resolutions, field of view and human natural motion are the key obstacles to accurate implementation. This study assesses the use of Generic Fourier Shape Descriptors (GFD) on person silhouettes for re-identification and furthermore identifies which sections of a subjects' silhouette is able to deliver optimum performance. Human silhouettes of 90 subjects from the CASIA dataset walking 0° and 90° to a fixed CCTV camera were used for the purpose of re-identification. Each subject's video sequence comprised between 10 and 50 frames. For both views, silhouettes were segmented into eight algorithmically-defined areas: head and neck, shoulders, upper 50%, lower 50%, upper 15%, middle 35%, lower 40% and whole body. A GFD was used independently on each segment at each angle. After extracting the GFD feature for each frame, a linear discriminant analysis (LDA) classifier was used to investigate re-identification accuracy rate, where 50% of each subject's frames were used for training and the other 50% were used for testing. The results show that 97% identification accuracy rate at the 10th rank is achieved by using GFD on the upper 50% segment of the human silhouette front (0°) side. For 90° images, using GFD on the upper 15% silhouette segment resulted in almost 98% accuracy rate at the 10th rank.