Person re-identification in multi-camera surveillance systems

Date of Award


Degree Name

M.S. in Electrical Engineering


Department of Electrical and Computer Engineering


Advisor: Vijayan K. Asari


In a system of cameras, it can be beneficial to track and identify people as they move through the scene. To solve this problem (called human re-identification) appearance matching through feature extraction must be applied to detected humans. We propose an algorithm that combines color features with soft biometric features; namely clothing identification that distinguishes pants from shorts, and long sleeve shirts from short sleeve shirts and backpack and sling bag detection. First, person extraction is performed using a mixture of Gaussians background subtraction and simple blob analysis. Next, a subject's color features are calculated using color histograms in the RGB and HSV color spaces. Clothing identification is performed by detecting skin on the top and bottom half of an extracted person. Backpack and sling bag detection is accomplished with advanced blob analysis. A label is calculated for each of these soft biometrics depending upon the percentage of frames a detection occurs in. Once a person is located, a signature of that real world subject is obtained by combining the labels and color features. An average signature of each individual that appears in a given camera (training camera) is then calculated by averaging all the signatures from a specific camera of that subject. The signatures gathered from other cameras (testing cameras) are then compared to the average signatures from the training camera using chi-squared distance measurements between the color histograms and between their cumulative distribution functions. These distances are ranked and then reranked according to the labels of the signatures. A correctly identified person has a ranking of one. These final ranks are shown in Cumulative Matching Characteristic (CMC) curves. The most significant of the many challenges involved are the variation of illumination conditions, pose, and viewpoint across the cameras. The algorithm is tested on the SAIVT-SoftBio dataset and promising results for human re-identification on multi-camera systems are observed. Research work is progressing for human re-identification based on computer rendered human models.


Biometric identification, Optical pattern recognition, Video surveillance, Electrical Engineering, Computer Engineering

Rights Statement

Copyright © 2015, author