Room Occupancy Detection from Environmental Sensors using Machine Learning

Smart buildings save energy by knowing whether rooms are occupied — HVAC and lighting follow people, not schedules. But cameras are intrusive and PIR sensors miss still occupants; environmental sensors (temperature, humidity, light, CO2) already present in HVAC systems carry the same signal passively. The UCI Occupancy Detection dataset (Candanedo & Feldheim) is the standard benchmark: minute-by-minute readings from an office, labeled occupied or not. This project trains and compares classifiers on the full dataset, and makes the learning tangible with a live demo — a logistic-regression model is actually fitted in your browser by gradient descent on genuine UCI measurements, and its predictions, learned weights and held-out metrics are all shown live. Because the demo trains for real and labels every figure as measured or design target, the student can defend the whole pipeline in the viva.

1. UCI occupancy data is loaded (date, temperature, humidity, light, CO2, humidity ratio, occupancy label); features are standardized and the train/test split follows the dataset's own files.
2. Logistic regression is trained by gradient descent, random forest and gradient boosting by their standard fitting procedures — all in one reproducible notebook.
3. The held-out test set is evaluated once: accuracy, precision, recall, F1, ROC-AUC and the confusion matrix are generated for each model.
4. In the web demo, a logistic-regression model is fitted live in the browser on an embedded UCI sample (204 rows, 300 epochs), with the loss curve drawn as it trains.
5. Slider readings are standardized with the training statistics, dotted with the learned weights, and passed through the sigmoid to give P(occupied | sensors).
6. The demo's measured metrics come from the embedded held-out sample (267 rows); the built-to-order report carries the full-dataset evaluation.

Tech stack:

• Python 3, scikit-learn (logistic regression, RF, GBM)
• NumPy, pandas, Matplotlib (metrics & plots)
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (live in-browser training demo)
• UCI Occupancy Detection dataset

Dataset & model details

• Dataset: UCI Occupancy Detection (Candanedo & Feldheim) — 20,560 training rows and 2,665 test rows of minute-level office sensor data. Features: temperature, relative humidity, light, CO2, humidity ratio. Label: room occupied or not.
• Task: Binary classification; input = 5-D environmental sensor vector, output = P(occupied | sensors).
• Model: Demo: logistic regression trained by gradient descent (300 epochs) in the browser. Built-to-order: logistic regression, random forest and gradient boosting trained on the full dataset and compared.
• Metrics: Demo (measured live on the embedded 267-row held-out sample): accuracy 97.4%, precision 93.3%, recall 100.0%, confusion TP 97 · TN 163 · FP 7 · FN 0. Built-to-order evaluation on the full test set reports accuracy, precision, recall, F1 and ROC-AUC per model. Full-dataset figures are design targets until the training run executes for the order.

• [Real in-browser training demo] A logistic-regression model is fitted live by gradient descent (300 epochs) on 204 genuine UCI training rows — the loss curve, learned weights and held-out metrics are computed in your session, not canned.
• [Live occupancy predictor] Move temperature, humidity, light, CO2 and humidity-ratio sliders; the probability comes from the fitted weights, with per-feature log-odds contributions shown.
• [Measured held-out metrics] Accuracy 97.4%, precision 93.3%, recall 100.0% on 267 held-out rows with the full confusion matrix — genuinely measured on the embedded sample in the demo.
• [Full-dataset training notebook] Logistic regression, random forest and gradient boosting trained and compared on all 20,560 UCI training rows with a reproducible evaluation on the 2,665-row test set.
• [Feature-importance analysis] The learned weights show light and CO2 dominating the decision — exactly what the occupancy literature reports — with the analysis documented.
• [Model comparison] The three classifiers compared on accuracy, precision, recall, F1 and ROC-AUC, with the tradeoffs (interpretability vs accuracy) discussed for the viva.
• [Occupancy presets] One-click occupied-room and empty-room sensor presets demonstrating how the model's decision boundary behaves.

• Complete training & evaluation Jupyter notebook on the full UCI dataset
• Trained classifier models with preprocessing and inference code
• Live in-browser training + predictor + model lab web demo
• Confusion matrices, ROC curves and model-comparison tables
• Project report PDF (sensor-fusion background, model theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (logistic regression, gradient descent, precision vs recall, sensor fusion, occupancy literature)