Publication
How Much Context Matters? A Comparison for Skeleton-Based Activity Recognition
Matthias Tschöpe; Björn Friedrich; Sizhen Bian; Paul Lukowicz
In: 2026 International Conference on Activity and Behavior Computing (ABC). International Conference on Activity and Behavior Computing (ABC-2026), Activity and Behavior Computing, March 9-12, Hakodate, Japan, IEEE, 2026.
Abstract
Automated and anonymized detection of normal and unusual behavior can support clinical staff in hospitals and care facilities, where continuous observation of patients is often not possible. Such systems can help to improve the safety of patients and employees by detecting unusual or potentially critical situations while protecting privacy by avoiding the use of raw video data.
In this work, we empirically compare models for classifying human activities using a 2D pose dataset recorded in a care-related context. We use five models and vary the temporal context by using six different sliding window configurations. In addition, we analyze how the use of a body-centered coordinate system changes the classification results. We evaluate all results using Leave-One-Subject-Out.
We focus on how temporal context, pose normalization, and the chosen models affect the classification results. The results show that graph-based and transformer-based models achieve similar classification results when sufficient temporal context is used. On our chosen dataset, the best classification results are achieved with ST-GCN by using a sliding window configuration of (180/90), we get an average accuracy of 79.65% and a macro F1-Score of 79.86%. Finally, we provide the GPU usage and power consumption for each model.