TinyML to Predict Occupancy Levels Using Microcontrollers

The ability to track occupancy levels of buildings such as campus attractions like the gym or dining hall is crucial to both users and administrators. An occupancy tracking system will enable users to effectively plan visits at optimal times to minimize wait times and equip administrators to allocate resources better. This work proposes a microcontroller-based system that captures WiFi probe requests at a location and uses a machine-learning model to determine the level of occupancy from the probe requests. The system runs on an ESP32-S3 and works by capturing WiFi probe requests, which WiFi-enabled devices such as phones and laptops emit, and uses a Random Forest regression model to predict the levels of occupancy. Occupancy data and WiFi probe requests in a gym area of 30m by 80m were collected and the data was used to train a Random Forest regression model. The system performance was evaluated when the regression model was deployed on the microcontroller and an off-device server. The system predicts the number of people in the gym with a mean absolute error of 1.743 off-device and 4.36 on-device. These results demonstrate that low-cost microcontrollers like the ESP32 are viable devices to run machine learning on microcontrollers (TinyML) and predict levels of occupancy using WiFi signals.