Development of Attention based Large-Scale Person Re Identification Systems

Abstract

Person re-identification (PRId) is critical in computer vision with significant surveillance, security, and public safety applications. This thesis explores the latest advancements in person re-identification techniques to address the challenges of matching individuals across different camera views and under varying conditions. The primary goal of this research is to enhance the accuracy and efficiency of PRId systems, ultimately contributing to improving security and surveillance systems. The thesis begins by providing an in-depth review of existing PRId methodologies, highlighting their strengths and limitations. It delves into the importance of feature extraction, metric learning, and deep neural networks in Re-ID systems. Furthermore, it discusses the challenges posed by variations in lighting, pose, occlusions, and camera viewpoints, which are inherent to real-world surveillance scenarios. A person-identification framework is proposed to highlight the within-part inconsistency problems that exist in the part-based PRId system, which assumes all the pixels in each part partition are homogeneous. Traditionally, hard and soft partition strategies were used to partition the body parts. However, the proposed method focuses on a recent convolutional part partition strategy. The proposed method demonstrates the effectiveness of convolutional part-partition over the hard and soft partition of body parts over three publicly available benchmark datasets. A graph-based PRId system is developed, which concentrates on two aspects: an explainable approach to attention selection and graph convolution methods. The proposed multi-channel framework utilizes visual features and attribute labels to represent each person uniquely. The proposed approach integrates an attention-based approach to evaluate the importance of different features by estimating the distance between the nodes. The research presents novel feature extraction and representation learning approaches to address these challenges, emphasizing deep convolutional neural networks (CNNs) and attention mechanisms. These techniques aim to capture discriminative information from images and encode it into compact and informative feature representations. Additionally, the thesis investigates the incorporation of domain adaptation and transfer learning to improve model generalization across different surveillance environments. Moreover, the thesis explores the utilization of large-scale PRId datasets and benchmark evaluation metrics to assess the performance of the proposed methods rigorously. It also discusses the ethical considerations and privacy implications associated with PRId technology, emphasizing the importance of responsible deployment. The experimental results demonstrate significant improvements in person re-identification accuracy and robustness compared to state-of-the-art techniques, validating the effectiveness of the proposed methodologies. Furthermore, the research discusses potential real-world applications of these advancements, including enhanced security in public spaces, more efficient search and retrieval in video archives, and improved human tracking in autonomous systems. In conclusion, this thesis contributes to the ongoing development of the PRId system by presenting innovative approaches to address its challenges. The research outcomes have the potential to reshape the landscape of security and surveillance systems, making them more reliable and capable of safeguarding public spaces effectively while respecting privacy concerns.