Mobile Re-Identification Based on Local Features Analysis

Abstract

Nowadays, there is a growing interest in security applications using embedded smart cameras. Despite the rising attention to re-identifying the people in a non-overlapping embedded camera network, there exist no comparative evaluation of the existing schemes, specially facial representations. Though, facial features offer the advantage of remaining stable over much larger time intervals in contrast to commonly exploited features for person re-identification, such as whole body appearance. However, the challenge in using faces for such applications, apart from low captured face resolutions, is that their appearance is largely influenced by changes in illumination, pose, viewpoint, background and partial occlusions. Therefore, in this paper, we present a comparative study of facial local features for the task of person re-identification using embedded smart cameras in a realistic surveillance scenario. The experiments are performed on a novel data set which we make publicly available for future research in this area. The contributions of this work are as follows: 1) evaluation of eight baseline facial local features for person re-identification using embedded smart cameras; 2) an empirical study of the performance of person re-identification systems based on local facial features when fused with auxiliary information such as face shape, face skin color tone, hair color, and cloth color; 3) comparative study of single-and multi-shot approaches; 4) a novel data set using embedded cameras for person re-identification with properties unavailable in exiting data sets.

Person Detection and Re-identification Across Multiple Images and Videos Obtained via Crowdsourcing

“…Then, we convert this region to the HSV color space to reduce the effect of illumination changes. Afterwards, we calculate the color histograms, and compare the histograms by using color correlation calculation in (1), where H1 and H2 are the two color histograms, N is the total number of histogram bins, and…”

“…Most of the existing work relies on the appearance-based similarity between images, such as color and texture of clothing, to establish correspondences. In general, recent approaches focus on three aspects; (1) designing subjectdiscriminative [30], descriptive and robust visual descriptors to characterize a person's appearance [1], (2) using feature transformation which projects features between different camera-dependent spaces, such as feature warping [19], and sparse basis expansion [18,16], and (3) learning suitable distance metrics that maximize the chance of a correct matching [2,14,23].…”