Deep vision-based occupancy counting: Experimental performance evaluation and implementation of ventilation control

Abstract

Recently, many researchers have become interested in occupant information and occupant-centric control (OCC) strategies, aiming for efficient building operation. Combining a camera with deep learning (hereafter referred to as deep vision-based occupancy counting) is a very effective method for occupancy counting, but there are not many studies evaluating its experimental performance. Furthermore, there are insufficient studies on implementing control using deep vision. The purpose of this study was to experimentally evaluate the performance of deep vision-based occupancy counting and to implement deep vision-based OCC in reality. First, we evaluated the performance of deep vision-based occupancy counting for six offices. Second, we implemented a deep vision-based energy recovery ventilator control strategy in a small office and compared the indoor air quality and energy consumption with those from traditional control strategies. As a result, deep vision-based occupancy counting showed significantly higher performance (root mean square error (RMSE): 0.883, normalized RMSE (NRMSE): 0.141). The larger the floor area, the more frequently the prediction of the number of occupants was lower than the actual number. The control results showed that deep vision-based ventilation control could properly maintain the indoor CO2 concentration with 24–35% lower ventilation rates compared to traditional ventilation control strategies. Furthermore, the proposed strategy was effective in reducing the electrical energy consumption of energy recovery ventilator and heat pump.