Chair: Gabriele Civitarese
Chair: Gabriele Civitarese
Keynote: LLMs and Physical Simulations for Sensor Based Context Recognition in Open Ended Settings
Paul Lukowicz (DFKI and University of Kaiserslautern, Germany)
Abstract: Generative AI is attracting rapidly growing interest in sensor-based human activity and context recognition. Yet, its impact in this domain remains far less transformative than in computer vision, where contemporary foundation models can reliably produce accurate descriptions for a wide range of arbitrary visual scenes. The prospect of an analogous capability for arbitrary "sensor scenes" is highly compelling, but, despite an expanding body of research and encouraging early results, it is still far from being realized in practice.
This talk will analyze the fundamental reasons behind the discrepancy between the remarkable progress in vision-centric generative AI and the comparatively modest gains observed in sensor-based context and activity recognition, including issues of the variability of the input space, data heterogeneity, limited large-scale labeled corpora, weak semantic grounding of sensor signals, and challenges in cross-modal generalization. Building on this analysis, the talk will outline how a combination of appropriately adapted multimodal generative models tightly coupled with physics-informed simulation and sensor-aware representation learning can solve those problems enabling more robust, generalizable, and scalable activity and context recognition systems.
Chair: Gabriele Civitarese
"Thoughtful Coordination": LLM Reasoning for Anomaly Detection in Complex Activity Sensing
Debjit Chatterjee and Sayantan Kuila (Indian Institute of Technology, Kharagpur, India); Sandip Chakraborty (Indian Institute of Technology Kharagpur, India); Soumyajit Chatterjee (Brave, United Kingdom (Great Britain) & University of Cambridge, United Kingdom (Great Britain))
Rotation Estimation Auxiliary Task for Wrist Inertial-Based Human Activity Recognition
Francisco Miguel Calatrava Nicolás (Örebro University, Sweden); Vitor Fortes Rey (DFKI, RPTU, Germany); Paul Lukowicz (DFKI and University of Kaiserslautern, Germany); Todor Stoyanov (Örebro University, Sweden); Oscar Martinez Mozos (University of Alicante, Spain)
Wear2Toy: Leveraging Wearable Devices for Hand-Held Objects Activity Recognition
Nicolò La Porta (Università Della Svizzera Italiana (USI) & University of Applied Science of Southern Switzerland (SUPSI), Switzerland); Carolina Di Maio (University of Applied Science of Southern Switzerland (SUPSI), Switzerland); Francesca D. Faraci (University of Applied Sciences and Arts of Southern Switzerland (SUPSI), Switzerland); Bronwyn Glaser and Gian Paolo Ramelli (Paediatric Institute of Southern Switzerland, EOC, Switzerland); Michela Papandrea (University of Applied Sciences of Southern Switzerland Switzerland (SUPSI), Switzerland)
IMU-based pre-impact, impact, and post-impact fall detection dataset
Muhammad Touqeer Ali (University of Verona, Italy); Cristian Turetta (Wenzhou Business College, China); Florenc Demrozi (University of Stavanger, Norway); Graziano Pravadelli (University of Verona, Italy)
Chair: Claudio Bettini
Causal Models Improve Reinforcement Learning for Pervasive and Robotic Tasks
Giovanni Briglia, Stefano Mariani and Franco Zambonelli (University of Modena and Reggio Emilia, Italy)
Privacy-Preserving Skeleton-Based Analysis of Visitor Behavior in Museums
Violeta Ana Luz Sosa-León (Universitat Jaume I de Castellón, Spain); Sergey Mukhametov (University of Muenster, Germany); Sergio Trilles (Universitat Jaume I, Spain); Joaquín Huerta (UJI, Spain); Joaquín Torres-Sospedra (Universitat de València, Spain)
Hierarchical Silent Command Recognition on Earables: A Multistage Learning Approach
Bambang Muharto (National Research and Innovation Agency, Indonesia); Egemen Işgüder and Ozlem Durmaz Incel (University of Twente, The Netherlands)
Chair: Gabriele Civitarese