room-occupancy-co2
One CO₂ sensor. Detecting occupancy without a camera, microphone or anything else.
See code →The problem
Room occupancy solutions that actually work (cameras, PIR sensors, people counters) cost hundreds of euros per room and generate sensitive data. How much can you do with a single €20 CO₂ sensor? That question has a physical limit: CO₂ measures gas concentration, not people. The challenge is knowing exactly where that limit is and extracting everything possible from it.
The data
Dataset from a real meeting room with CO₂, temperature and humidity sensor at 1-minute resolution. Occupancy labels (0/1) collected independently from the CO₂ signal. The dataset covers different usage patterns: short meetings, long sessions, occasional ventilation.
Approach
- Reframed as a latent state problem: occupancy is not directly observed, CO₂ is a noisy proxy with inertia — that is an HMM, not a direct classifier
- Viterbi algorithm on the HMM: finds the most likely sequence of states (occupied/free) given the sequence of CO₂ observations
- Explicit inertia modelling: CO₂ takes 5 to 15 minutes to reflect an occupancy change — the model knows this and does not react to every spike
- Hysteresis on transition thresholds to avoid oscillating states from sensor noise
- Comparison with direct threshold classification: the HMM wins on transitions, loses on reaction speed — both options are documented
Tools
python
jupyterResult
Reliable binary occupancy detection with a single low-cost sensor. The most valuable result is not the accuracy: it is the explicit analysis of the physical limit. Trying to predict how many people are in a room from a CO₂ sensor is a category error — the gas does not carry that information. The project proves it with data.
Next steps
The direct extension is adding a second sensor (temperature or VOC) to improve detection speed on transitions — CO₂ alone is slow. There is also room to explore whether the model generalises to other rooms without retraining, or whether it needs per-room calibration due to ventilation and volume differences.