Application of vision-based occupancy counting method using deep learning and performance analysis

Recently, several researchers have attempted to detect occupancy information and use it to reduce building energy. Although occupancy counting using a camera and deep learning is a very effective method that has recently emerged, there are few cases in which the experimental performance and utilization have been comprehensively evaluated. The purpose of this study is to comprehensively evaluate the applicability of vision-based occupancy counting using the latest deep learning model. First, this study experimentally evaluated the performance of the vision-based occupancy-counting method in two offices and investigated the user's acceptance of this method using a questionnaire. Second, the energy-saving performance of various occupancy-centric control strategies applied to HVAC and lighting was analyzed. Experimental results showed high performance for a small office of fewer than five people (NRMSE: 0.0435) and lower performance in a larger office (NRMSE: 0.0918). In addition, despite several occupants feeling privacy invasion by the camera, they responded that the system could be accommodated to reduce building energy. Simulation results indicated that occupancy-centric control using the number of occupants could reduce annual HVAC and lighting energy in small offices by 10.2%. In addition, among several occupancy-centric control strategies, modulating the outdoor airflow rate was found to be the most effective in saving energy.