Introduction
Diagnosing and managing Alzheimer’s disease relies heavily on assessing a patient’s ability to perform Instrumental Activities of Daily Living (IADLs)—tasks like preparing a meal, managing medication, or using a telephone. Traditionally, these assessments are subjective, based on caregiver reports or brief clinical observations.
Researchers are now turning to "egocentric" (first-person) vision to provide clinicians with an objective view of how patients navigate these tasks in their own environments. While wearable cameras offer a privileged viewpoint, they introduce two major hurdles: the sheer volume of data produced and the sensitive nature of recording inside a patient's home. This paper explores technical modules designed to automate video analysis and provide actionable data while maintaining privacy.
Methodology and Hardware
The study utilized a lightweight, non-intrusive hardware setup: a GoPro camera (recording at 1280x960p, 30fps) mounted on a specialized shoulder-worn jacket. This placement ensures the camera captures the patient's workspace and hand movements without the discomfort or instability of head-mounted devices.
Technical Modules
The proposed system focuses on two primary technical areas to transform raw video into clinical insights:
1. 3D Location Estimation
To understand where activities occur, the team employs a "Structure from Motion" (SfM) pipeline. By detecting and matching interest points across sequential video frames, the system reconstructs a 3D point-cloud of the home environment. This allows the software to track the patient’s precise 3D position even when traditional ambient cameras would be obstructed by furniture or caregivers.
2. Visual Saliency Modeling
This module predicts which parts of a scene are most important to the patient. The researchers compared the gaze of the Actor (the patient) with the Viewer (the clinician). Using Gaussian filtering, they converted raw eye-tracking data into "heatmaps" of attention. A key finding was the "Action-Gaze" time shift: physiological studies show a person’s gaze typically anticipates their motor actions by about 500ms.
Experimental Results
The study utilized the GTEA (Georgia Tech Egocentric Activities) dataset for saliency validation and Dem@care Project data for 3D environment testing.
- Temporal Alignment: Quantitative analysis using AUC (Area Under the Curve) metrics showed that the correspondence between the actor's and viewer's attention peaks when the viewer’s data is shifted forward by approximately 10 frames (roughly 667ms). This suggests that clinicians naturally "catch up" to the patient's intent.
- Mapping Accuracy: The SfM pipeline successfully generated high-fidelity 3D models of laboratory environments, proving that egocentric video can reliably determine a patient's position. This spatial context significantly narrows down the possible activities a computer vision system needs to recognize (e.g., kitchen coordinates imply meal preparation).
Conclusion and Future Work
While the system demonstrates high potential for objective IADL assessment, several challenges remain. Future work will focus on refining automated privacy filters to "summarize" video without exposing sensitive personal information. Additionally, the team plans to move beyond controlled lab environments to test the system's robustness against the clutter and complexity of real-world home settings. The ultimate goal is to integrate these modules into the broader Dem@care framework, providing a fully automatic feedback loop for both clinicians and patients.
Summary updated