Sensing social interactions in buildings using ambient sensors: A chamber study

Buildings bear the dual responsibility of achieving environmental sustainability and supporting social value. Accurately sensing indoor social interactions, particularly verbal conversations, is key to quantifying spatial social vibrancy. This study proposes a strategy for sensing social interactions using non-intrusive and privacy-preserving indoor environmental quality sensors, validated through an empirical study conducted in a controlled climate chamber. We deployed 27 sensors to monitor 11 categories of physical indicators and systematically varied occupant density, activity types, and air change rates (ACH 0, 2, and 4.39/h). We established a three-layer feature extraction framework comprising universal baseline, indicator-specific, and spatial features and compared the performance of linear Logistic Regression and non-linear LightGBM models for predicting interactions. Experimental results indicate that the sound level serves as the main ambient indicator for social recognition, achieving near-perfect classification accuracy. Environmental indicators such as CO2, temperature, and TVOC demonstrate high performance in occupancy detection, but reduced performance in interaction detection. Spatial analysis reveals a “spatiotemporal resolution coupling” pattern: high-frequency monitoring relies heavily on near-field sensing, whereas low-frequency monitoring depends predominantly on spatially distributed comparisons. Furthermore, multimodal fusion strategies demonstrate potential to mitigate the limitations of single non-acoustic indicators. Ultimately, by aligning sensor layouts with sensing frequency and leveraging multimodal data, ambient sensing serves as a powerful, privacy-preserving tool for capturing social interactions in buildings.