Thesis Project Form

Sleep disorders such as sleep-related hypermotor epilepsy (SHE), disorders of arousal (DOA), and REM-sleep behaviour disorder (RBD) significantly affect people' health, quality of life, and the well-being of their families, and their diagnosis still relies heavily on the manual review of long polysomnographic and video-EEG recordings. In this context, video-based monitoring offers a promising, non-invasive route to capture motor and behavioural patterns over long, naturalistic observation windows: the present thesis focuses on nocturnal monitoring. Recent work on the project has produced deep-learning models capable of classifying the type of a nocturnal motor event from a pre-selected video clip; however, the step before classification, automatically locating when and where such events occur within hours of mostly uneventful recording, and distinguishing them from the movements of nurses and caregivers in the room, remains an open challenge and is the prerequisite for any fully automated monitoring of whole-night recordings.

The long-term goal of this project is the development of an automatic monitoring pipeline able to process whole-night infrared recordings and forward to a downstream classifier only the segments containing clinically relevant motor activity. In order to accomplish this goal, the proposed thesis has different aims:
the implementation and evaluation of motion-localization techniques (e.g. optical-flow–based methods) robust to the infrared imaging and lighting variations typical of the recording setting;
the integration of person detection and multi-object tracking to identify and follow each individual in the scene, anchoring the patient's identity to the bed region and using region overlap to flag caregiver interventions;
the design of a decision logic that attributes movement to the patient rather than to other people present;
the development of a binary screening stage that separates pathological from physiological movements, with an uncertainty-based mechanism that routes ambiguous cases to human review;
the experimental validation of the pipeline on the existing annotated dataset, with a feasibility demonstration on whole-night recordings.

The student will learn:
1. Computer vision techniques to process and analyze images and long-form videos; 2. Deep-learning–based object detection methods to localize persons in the scene; 3. Multi-object tracking algorithms to maintain persistent identities across frames;
4. Optical-flow and motion-analysis methods for the localization of events in video; 5. Machine-learning techniques for binary classification with uncertainty estimation; 6. The design of evaluation protocols for clinical and imbalanced data;
7. Improve the knowledge of Python and of the main libraries for video processing and deep learning (e.g. OpenCV, PyTorch).