Most works conceal people’s visual privacy by using blurring or pixelating effects to modify an image. In a privacy-by-context approach, a level-based visualisation scheme to protect privacy is proposed. Each level establishes the way in which the video images are modified and displayed and, therefore, the provided protection degree. In this scheme, the appropriate level is dynamically selected according to the context, therefore modifying a non-protected image before it is displayed. The context has to provide enough information in order to empower people to adapt privacy to their preferences, in such a way that they can decide by whom, how and when they can be watched. The context is modelled by different variables: (i) the observer; (ii) the identity of the person (to retrieve the privacy profile); (iii) the closeness between the person and observer (e.g., relative, doctor or acquaintance); (iv) appearance (dressed?); (v) location (e.g., kitchen); and (vi) ongoing activity or detected event (e.g., cooking, watching TV, fall). Therefore, an accurate recognition of the context is paramount to provide the appropriate privacy level. This project will investigate techniques to recognise accurately these variables and it will validate them under different use case scenarios.
Start date: April 2021
Expected end date: April 2024
The first variable to be addressed in this project has been appearance recognition and nudity detection in private spaces. During the first months, a literature review on nudity recognition has been carried out. The next step was to develop machine learning methods for nudity recognition using deep learning. However, the main problem with these methods is that the require large amounts of data and the available datasets were either small or low quality. Therefore, by using datasets for homogenous tasks such as garment recognition, a new skin dataset has been created as an extension of the FashionPedia garment and clothing dataset. In collaboration with ESR1, who has studied nudity from the social science discipline, a study over focus group has been done and then a methodology for nudity level recognition has been developed based on that study. Recently, using this knowledge, ESR14 has developed a deep learning method for skin segmentation integrating state-of-the-art semantic segmentation and attention models.
Weakly supervised human skin segmentation using guidance attention mechanisms
Multimedia Tools and Applications, 2023
ODIN: An OmniDirectional INdoor dataset capturing Activities of Daily Living from multiple synchronized modalities
2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Vancouver, BC, Canada, 2023, pp. 6488-6497
A Fallen Person Detector with a Privacy-Preserving Edge-AI Camera
In Proceedings of the 9th International Conference on Information and Communication Technologies for Ageing Well and e-Health ICT4AWE - Volume 1, 262-269, Prague, Czech Republic, 2023
Underneath Your Clothes: A Social and Technological Perspective on Nudity in The Context of AAL Technology
In Proceedings of the 15th International Conference on PErvasive Technologies Related to Assistive Environments, pp. 439-445, 2022.
From Garment to Skin: The visuAAL Skin Segmentation Dataset
In International Conference on Image Analysis and Processing. Springer, Cham, pp. 59-70, 2022.
Kooshan obtained his bachelor degree in Electrical Engineering (2015) from KNTU and a master’s degree focused on Signal Processing (2018) from Iran Broadcasting University. During his studies, he mainly explored the field of machine learning and after graduation, he worked as a Computer Vision engineer in different startups in Iran and developed large-scale machine learning services.
Kooshan Hashemifard
University of Alicante
Department of Computing Technology
Ctra. San Vicente del Raspeig, S/N
03690 San Vicente del Raspeig, Spain
Email address: k.hashemifard@ua.es