A1344
Title: Robust spatio-temporal graph modeling for behavior monitoring under heterogeneous environments
Authors: Jiwon Im - Chungnam National University (Korea, South) [presenting]
Yunchae Jung - Chungnam National University (Korea, South)
Minsu Park - Chungnam National University (Korea, South)
Abstract: Patient distress detection in clinical facilities necessitates high operational reliability to ensure patient safety. Conventional skeleton-based methods using graph convolutional networks often encounter performance bottlenecks in low-light conditions, where poor illumination induces significant noise in pose estimation and destabilizes joint detection. A real-time detection framework is developed to maintain robustness against adverse illumination by integrating skeleton-based motion analysis with scene-level context. Joint coordinates extracted from video frames are structured as spatio-temporal graphs, encoding spatial relationships and capturing dynamic motion patterns. Integrating scene-level context features complements the skeletal representation, enhancing situational inference. The challenge of performance degradation in low-light settings is addressed through an illumination-adaptive training strategy utilizing both standard and synthetically darkened data generated via gamma correction and Gaussian noise. This strategy stabilizes joint extraction and ensures reliable classification without computationally intensive image enhancement. Experimental results demonstrate that the framework consistently distinguishes routine actions from distress indicators, maintaining performance under low-light comparable to normal illumination. The achievement of high-frequency processing speeds confirms the feasibility for real-time deployment in clinical monitoring systems.
