Privacy-preserving Person Re-Identification in Event-based Vision

In modern cities, extensive surveillance camera systems are installed to enhance public safety and security, generating large volumes of visual data. The automatic processing and interpretation of this data drive the advancement and use of visual data analysis technologies. One of the important research topics within surveillance systems is person re-identification (ReId), which aims to recognize the same person across multiple non-overlapping camera views. Re-id has gained significant interest in the computer vision community as an enabling technology for intelligent visual surveillance systems, e.g., tracking in non-overlapping views, and forensic and security applications. Nevertheless, for visual surveillance, collecting images and videos with always-connected vision sensors in public spaces puts individual privacy at stake while increasing resource consumption. In this thesis, we first investigate a recent research problem in classical ReId., "semantically aligned person ReId." We proposed an aligned person re-id model that utilizes human parsing and object detection networks to obtain the pixel-level part-aligned representations for person re-id. Afterward, the main focus and goal of the thesis was to address the privacy-related issues in video surveillance (e.g., person ReId) by employing Neuromorphic vision sensors (event cameras). As a result, we developed a novel end-to-end privacy-preserving person ReId system in event-based vision. Additionally, we proposed a first-ever person ReId dataset captured with event cameras to address the unavailability of event-based datasets.