Resolution Representation Based Person Re-Identification for Smart Cities Using Deep Neural Networks (DNNs)

Abstract

The advancement of deep learning has facilitated rapid progress in person re-identification (re-id) task. Its applications in intelligent video surveillance made it a key component of today's smart cities infrastructure. Person re-id task is aimed to identify the person in distributed camera setup with non-overlapping views. The feature extraction process is an important part of person re-id technique. The present state of art methods mostly used ResNet as a backbone for feature extraction, which results in low geometric transformation modeling and low-resolution representation learning. We addressed these two major challenges by integrating deformable convolution module to enhance the transformation modeling capability and replaced the traditional ResNet backbone for person re-id with the novel feature extraction network named as HRNet, which is based on high-resolution representation learning without any additional supervision. The verification of our approach performance is done by conducting an experiment on a person re-id dataset named Market-1501. We achieved 90.57% Rank-1 accuracy and 75.43% mAP, outperforming the ResNet baseline results, which confirmed the effectiveness of our approach and will have a promising future in person re-id.