A preliminary analysis on the use of low-cost data streams for occupant-count estimation

This paper presents an analysis of occupancy and occupancy-related data gathered from an academic office building. The data set contains records from the WiFi access points, motion detectors, CO2 sensors, light power and plug-load meters, and camera-based image processing sensors. Concurrent ground-truth occupant counts were collected on five days. Two sensorfusion model formalisms were developed to blend the information in individual data streams: multiple linear regression and artificial neural networks (ANNs). The results indicate that low-cost data streams that are not intended for occupancy sensing, such as WiFi traffic, CO2 concentration, and light power and plug-load data, perform at least as accurately as motion detectors and camera-based image processing sensors in estimating the total number of building occupants.