CO2 based room occupancy detection : an IoT and machine learning application

A CO2 based detection occupancy solution is of great use in smart buildings which automatically control heating, air conditioning,... with help of sensors. An empty building should lower its heating if it detects a pattern of non-occupancy or raise an alarm if it finds an abnormal behaviour. At first, we want to record environmental variables gathered by IoT sensors. Then, build a complete data set of samples composed of co2 ppm, temperature, humidity, light and the actual number of people occupying a room in order to have a solid training set suitable for supervised machine learning. We will then use machine learning algorithms to retrieve more useful and meaningful information such as deducing the precise number of people in a room based on CO2, temperature, humidity,... Those can then be valuable to reduce energy consumption, find patterns of occupancy in buildings and take actions based on those results. We also want to extend our work to other usages than room occupancy and find similar possible deployments. In this work, we will present how to gather valuable data from IoT sensor, semi-manual systems and a raspberry pi equipped with a camera, how we pre-processed the data and feature engineered the samples. We then evaluate various machine learning techniques and combine them in order to predict the exact room occupancy. We estimate the number of occupants using a hard voting technique combining Random Forest, K-Nearest Neighors and Multi-Layer Perceptron classifiers and regressors. While training the model and testing it against samples from the same room, it can achieve a strict (with no tolerance) accuracy of 85%. However, we also found that while trained on several rooms and tested on the same set of places, the method can actually keep its accuracy. We also found empirically that the most valuable features for this task are the CO2 level, the temperature, the humidity and the light of a room.