Neural Architectures and Approaches for Person Re-Identification in Autonomous Surveillance System

Abstract

Video-based person re-identification is the task of recognizing the same person across different cameras and video clips. It is an important part of automated surveillance systems because a person may disappear from one camera, become partially blocked, or reappear later in another camera view. This task is difficult because people can look different when the camera angle, lighting, distance, background, or body pose changes.

This thesis studies how video-based person re-identification can be improved under three practical deployment conditions. First, it presents a resource-aware method for cases where only a small amount of labelled data is available. The method begins with one labelled video clip per person and gradually adds reliable, automatically labelled examples to improve training while keeping the model efficient. Second, it develops a fully supervised method for cases where labelled training data are available. This model uses information from the entire video sequence as well as local body-region details to improve recognition under occlusion, pose changes, and background clutter. Third, it introduces a transfer-based method for cases where a model trained on one camera network must be used in another network without new manual labels. This method helps the model recognize people more reliably when the camera setup, viewing angle, lighting, or background changes.

Experiments on several video-based person re-identification datasets show that the proposed methods improve performance under different levels of supervision, computational cost, and camera variation. Overall, the thesis provides a practical study of how video-based person re-identification systems can be designed for label-scarce, fully labelled, and cross-camera deployment settings.