Sensing Image Regions for Enhancing Accuracy in People Re-identification

Abstract

Video surveillance systems are widely used in the public and private sectors for maintaining security and healthcare purposes. Performance of surveillance systems directly depends on their accuracy in re-identification. There are three regions in a camera view, including person’s body, background, and possible carried object by the person. Background, in existing approaches, is either overlooked or treated like a person’s body in re-identification. In this paper, these three regions are considered in re-identification but with different importance. In our proposed technique, first, the input image is semantically segmented into the three regions using a deep semantic segmentation approach. Then, the effect of each region on characteristic features of people is tuned depending on the region’s importance in re-identification. The proposed technique, leveraging robust descriptors, such as the Gaussian of Gaussian (GOG) and Hierarchical Gaussian Descriptors (HGD), can enhance existing methods in dealing with the challenging issues such as partial occlusion caused by carried objects and background in re-identification. Experimental results on commonly used people re-identification datasets demonstrate effectiveness of the proposed technique in improving performance of existing re-identification methods.