A new algorithm for real-time detection of window opening area in residential buildings

Abstract

The variation in window opening areas by individuals can lead to significant differences in air change rates, indirectly affecting indoor air quality and building energy consumption. While existing research primarily focuses on whether windows are open or closed due to data limitations, this study proposes a novel detection method for window opening areas based on the one-stage object detector You Only Look Once version 5 (YOLOv5). Two sub-models, YOLOv5s and YOLOv5l, were evaluated using training datasets of various sizes, and their performance was validated with test datasets and a video. The YOLOv5l_927 model emerged as the most generalizable, achieving a recognition accuracy exceeding 91% in both the training and test sets, outperforming the Faster R–CNN model by 12%–15%. Moreover, the YOLOv5l_927 model demonstrated real-time detection capabilities in a video, with an inference speed of 89 frames per second. The application of this method to monitor nine bedrooms revealed that 50% of the window opening operations involved adjustments between different opening areas. Occupants were more inclined to open the window area to 75–100%. This innovative detection method enhances the accuracy of building energy simulations by refining window opening behavior based on opening area.