Estimation of indoor occupancy level based on machine learning and multimodal environmental data

The detection of presence and occupancy has been identified as an important research topic in the context of anomaly detection and energy efficiency. In our paper, we compare the performance of four different classifiers (Support Vector Machine, Random Forest, AdaBoost, Gradient Boosting) in the estimation of the indoor occupancy level based on raw ambient sensor data. We compared three ensemble classes (AdaBoost, Gradient Boosting and Random Forest) and the Support Vector Machine (SVM) to classify the level of occupancy in two enclosed spaces in Monterrey, Nuevo León, Mexico, mentioned in a publicly available data set. Data were recorded in two spaces: a fitness center between 18 September and 2 October 2019 (Gym data set) and a living room in a private residence between 14 May and 4 June 2020 (Home data set) each second. The estimation of room occupancy was carried out as classification (assignment of one of the four occupancy classes) using ensemble classifiers and SVM (support vector machine). The highest performance measures were achieved in the Gradient boost (a 0.9540 recall, a 0.9571 accuracy, a 0.9834 accuracy, a 0.9555 home data set, and a 0.9996) and the F1 scores of a 0.9994 gym data set. The lowest overall performance was observed for Random Forest (0.1979 recall, 0.1752 precision, 0.1660 accuracy, and 0.1819 F1 score). The results indicate that Gradient Boosting is the most suitable method for estimating indoor occupancy based on raw environmental data, whereas SVM is the least suitable method, regardless of the applied kernel.