Graduate Student Seminar Series – Seyed Pourya Moghadam Kouhi

Graduate Student Seminar Series
Please ensure you invite your Principal Investigator by adding their email via the ‘Add Guest’ button and they will also be notified of your presentation.
Location: MS2158 – 1 King’s College Circle
Presentation Title: Multimodal AI Systems for Non-Visual Activity of Daily Living Monitoring in Smart Home
Abstract:
Monitoring Activities of Daily Living (ADLs) is fundamental to the development of intelligent healthcare technologies that support aging in place and independent living by characterizing behavioral patterns, functional ability, and early indicators of health decline. Continuous monitoring of daily routines can further support the longitudinal assessment of functional ability and early signs of cognitive decline in older adults living independently. Despite significant progress, ADL monitoring systems still face a trade-off between accuracy, computational complexity, and the need for continuous video recording.
To address this gap, this thesis proposes approaches to enable accurate, non-visual, and lightweight multimodal ADL monitoring systems suitable for edge deployment. The contributions are organized around three interconnected research components.
First, we propose a novel lightweight fusion-based architecture, called EFACT (Attention-Based Cross-Modal Temporal Modeling for Fine-Grained Activity Recognition Using RGB-D Data), designed for fine-grained human activity recognition using RGB and depth modalities. Using the Toyota Smarthome Dataset, we demonstrate that EFACT outperforms state-of-the-art approaches under the established benchmark protocols for held-out participants and camera viewpoints. Additionally, we show that the integration of quantization techniques (PTQ and QAT) allows EFACT to operate efficiently on resource-constrained platforms.
The second contribution of this thesis is the exploration of Non-Intrusive Load Monitoring (NILM), or estimating appliance-specific consumption from aggregate smart meter signals, as a building block for non-visual ADL recognition. We propose EdgeTCNformer, a lightweight causal framework for appliance-level power and state estimation that supports efficient local inference on residential edge gateways. Evaluation on four diverse datasets demonstrates a favorable accuracy–computational-efficiency trade-off under repeated sample-level and leave-one-house-out protocols, while zero-shot cross-dataset evaluation exposes substantial limitations under regional and acquisition-domain shift.
The third contribution of this thesis is E2R-FuseNet, an ADL recognition system for residential settings that combines NILM with millimeter-wave (mmWave) radar. We demonstrate that E2R-FuseNet outperforms radar-only systems in terms of accuracy and achieves performance comparable to some state-of-the-art vision-based systems, while offering computational advantages and eliminating the need for continuous video recording during inference.
Overall, the contributions of this thesis collectively advance the development of next-generation intelligent home monitoring technologies, supporting independent living and enabling scalable, non-visual healthcare solutions.
Supervisor Name: Dr. Azadeh Kushki
Year of Study: 3
Program of Study: PhD
Powered by Calendly.com