Commercial activity trackers frequently fail to meet the needs of older adults by focusing on generic metrics (e.g., step counts) while overlooking the personally meaningful activities that they may want to track in their everyday life. This disconnect stems from a fundamental challenge: the difficulty of capturing accurate ground-truth labels for complex, real-world activities, which forces a reliance on generic, one-size-fits-all models. My doctoral research seeks to address this problem by developing a human-centered approach to recognize activities that truly matter to individual older adults, moving beyond basic posture detection to a rich, semantic understanding of daily life. My research will culminate in a framework for a "teachable" activity tracker that older adults can train themselves. Part of my completed doctoral work establishes the need for personalization by demonstrating significant variability in older adults' movement patterns and deconstructs the challenges of collecting data "in the wild," quantifying the costs and benefits of triangulating sensor data with user-provided verbal reports. This is complemented by technical explorations into data-driven methods for optimizing on-body sensor placement. Ongoing work investigates the feasibility of older adults training their own personalized models in-situ through transfer learning. We are going to explore models capable of inferring high-level, meaningful activities (e.g., "gardening," "playing golf") from a stream of low-level sensor data. Ultimately, this thesis contributes initial steps toward a reframing of activity recognition—from an emphasis on technological capability to an emphasis on user-valued activities—thereby offering insights that could support the development of more personalized tools for older adults.
Hossein Khayami (Thu,) studied this question.