ESR13 - Siddharth Ravi

ESR13’s research focused on privacy preservation in video-based AAL applications. His work addresses a major concern in AAL: how to provide meaningful support through video monitoring while safeguarding users’ visual privacy. His research proposes an end-to-end Privacy by Design framework tailored for omnidirectional (fisheye) cameras with top-down (zenithal) views, commonly used in indoor environments like smart homes.

The motivation behind his research stems from the increasing use of video technologies to monitor older adults and people with disabilities. While video allows for rich contextual understanding, it also introduces serious privacy concerns—especially in intimate settings such as bathrooms or bedrooms. ESR13 draws upon Nissenbaum’s theory of Contextual Integrity and Privacy by Design principles to build systems that respect users’ expectations around how visual data is collected, processed, and shared.

One of ESR13’s major contributions is the development of a new taxonomy of Visual Privacy Enhancing Technologies, categorising techniques like anonymisation, obfuscation, encryption, and data minimisation. He links these techniques with high-level system design principles, helping to bridge the gap between technical and legal/ethical approaches to privacy.

As there is a lack of datasets containing videos from such omnidirectional cameras to train machine learning models for the task, ESR13 acquired ODIN (OmniDirectional Indoor dataset), a large-scale multi-modal dataset capturing real-world activities of daily living from top-view omnidirectional cameras. ODIN includes over 300,000 images and additional synchronised data: lateral-view RGB and depth videos, egocentric footage, physiological readings, and 3D scans of home environments. Building on this dataset, ESR13 is adapting an existing privacy-preserving visual pipeline—originally built for conventional RGB images—to work with zenithal-view omnidirectional data. The ODIN dataset has been employed during the project for various scene understanding tasks, especially human semantic segmentation and human pose estimation from top-view images. Various labels for scene understanding tasks, including for semantic segmentation and pose estimation models, have been created using semi-automated pipelines and manually verified. Using this, the first instances of machine learning models aimed at pose estimation and real-time human semantic segmentation on ceiling-mounted omnidirectional cameras have also been trained.

In addition to the technical developments, ESR13 is exploring the legal compliance of his privacy pipeline, particularly with the EU’s General Data Protection Regulation (GDPR) and the AI Act. This includes evaluating how different design choices align with consent, data minimisation, and user transparency obligations.

Alongside this, in a collaboration with ESR11, a study of the fairness of common visual privacy preservation algorithms was also conducted and published.