Activity Recognition using Multi-Class Classification inside an Educational Building

Activity Recognition can be referred to as the process of describing and classifying actions, pinpoint specific movements, and extract unique patterns from the dataset using heterogeneous sensing modalities. Activity Recognition approaches have garnered the attention of researchers in the energy management domain to enhance energy utilization in buildings. In our experiment, we define activities as a combination of different actions, which are detected using multiple sensors. To learn insights for the various activities, we used inexpensive Passive Infrared (PIR) sensors in the test-bed. This study aims at gaining high-level knowledge about activities from the low-resolution sensors deployed. For accurate occupancy counts, we have used 3D Stereo Vision Cameras at the entrance, and count lines are defined to capture the transitions of inflow and outflow of multiple occupants. Multi-class labels enable activity recognition on the collected dataset. The multi-class labels used are 1) Moving, 2) Stagnant, 3) Outside, 4) Both (Moving and Stagnant), 5) No activity inside. The labeling for the multiclass is done through an algorithm using supervised learning. The data acquisition gets carried out from 23rd November to 3rd December 2018, spanning over a period for 11 days. The results document that Gradient Boosting Classifier outperforms any other Machine Learning Classification (MLC) algorithm with an accuracy of 97.59% and an F1 score of 97.40% for activity recognition. This paper also explicitly highlights the challenges and limitations faced during the initial phase for the deployment, and it identifies the key research trends and directs towards the potential improvements in the field of occupancy sensing for energy-efficient buildings.