FRAMEWORK FOR PERSON RE-IDENTIFICATION

Abstract

The field of surveillance and forensics research is currently shifting focus and is now showing an ever increasing interest in the task of people re-identification. It is a fundamental task in the automated surveillance system essential to track a person in multi-camera setting. Re-identification (Re-ID) can be defined as a process of identifying the resemblances of a set of probe images or a single probe image representing a single person from a set of gallery images of people taken from the same or different cameras placed at different locations. However, established identification techniques being used presently face many difficulties and shortcomings. Traditional surveillance cameras provide low resolution images and thus state of the art face recognition and iris recognition algorithms cannot be easily applied to surveillance videos and images as people are required to face the camera at a close range. The different lighting environment inherited by each camera scene and the strong variations in illumination induce large changes in their appearance of a person walking through the scene. In addition, people images are occluded by other passers-by or objects in the scene making people detection further difficult to achieve. So to address the challenges in person re-identification problem, major contributions are being made to design robust feature representations and good discriminant metrics to evaluate the similarity between two person images effectively. Recently, it seems that more researches have been made in metric learning. So in this work potentials of feature design are emphasized. A novel and efficient person descriptor is proposed by utilizing knowledge on dense sampling of low-level statistics. We simply model a multi-layer representation of pixel features using multiple Gaussian distributions. More specifically, first we have created a pixel feature representation using color and texture information, then we have extracted local patch Gaussians inside overlapping regions. Further to describe these regions we have again used a Gaussian model on the local patch Gaussians. Thus we are able to use the discriminative properties of mean and covariance together along with considering the local structures in the image while globally analyzing them. For metric learning we have used a discriminant subspace and kernel learning method described in Liao et. al (2015), i.e. XQDA. It learns a discriminant low dimensional subspace and a QDA metric on the projected subspace, simultaneously. The proposed descriptor and metric is compared on benchmark datasets with current state-of-art-methods and has proven to give efficient and robust results. It exhibits remarkably high performance which outperforms some of the state-of-the-art descriptors for person re-identification. We have then considered it as a retrieval or recognition problem with the expectation that the n-highest ranked matches in the gallery will provide an identity for the unknown person, thereby identifying the probe.