Human Activity Recognition using Smartphone Sensors

Every smartphone carries an inertial measurement unit — a tri-axial accelerometer and gyroscope sampling motion dozens of times per second. That stream is rich enough to tell walking from sitting, stairs from flat ground, without any camera or GPS: the signature of each activity lives in the rhythm and amplitude of the signal. Turning raw sensor traces into activity labels is the human activity recognition (HAR) problem, and it matters wherever a camera is impractical or intrusive — fall detection for the elderly, step counting that does not drain GPS, context-aware phones that silence themselves in meetings. The standard teaching benchmark is the UCI HAR dataset (Anguita et al., ESANN 2013): 30 volunteers aged 19–48 wore a Galaxy S II on their waist while walking, climbing stairs, sitting, standing and lying down, producing 10,299 labeled 2.56-second windows at 50 Hz. This project trains a two-layer LSTM directly on those windows, evaluates it on nine held-out subjects the model never trained on, and ships a sensor console demo with live animated streams, per-activity waveform patterns, and a per-class evaluation view. The report is upfront about the classic confusion pair — sitting versus standing — and about the fixed waist placement the dataset assumes.

1. Raw 50 Hz accelerometer and gyroscope signals are noise-filtered; gravity is separated from body acceleration with a 0.3 Hz Butterworth low-pass filter.
2. Signals are segmented into 2.56-second windows (128 readings) with 50% overlap, each window labeled with one of six activities.
3. Windows are split by subject — 21 subjects for training, 9 for testing — so evaluation measures generalization to new people.
4. The LSTM reads each 128-step window (6 channels: 3 accel + 3 gyro axes) and the final hidden state feeds a 6-way softmax.
5. Categorical cross-entropy is minimized with Adam; validation F1 on held-out subjects drives early stopping.
6. The console demo streams simulated sensor data through the same windowing and displays the predicted activity with confidence.

Tech stack:

• Python 3.10, TensorFlow/Keras (two-layer LSTM, training loop)
• NumPy, pandas (signal loading, windowing, 561-feature handling)
• SciPy (Butterworth filtering, following the dataset's preprocessing)
• scikit-learn (SVM baseline, metrics, confusion matrix)
• Matplotlib, Seaborn (waveform plots, training curves, confusion heatmap)
• Jupyter notebook (buyer-run training and evaluation)
• Flask demo app with the animated sensor console
• UCI HAR dataset (Anguita et al., ESANN 2013 — 30 subjects, 6 activities, 10,299 windows)

• [Two-layer LSTM classifier] 128-unit LSTM layers over 128-sample windows of accelerometer and gyroscope signals; dense softmax over the six activities.
• [UCI HAR data pipeline] Loader for the 10,299-window dataset with the official subject-disjoint split: 21 subjects train, 9 subjects test.
• [Live sensor console] Animated tri-axial accelerometer and gyroscope streams with activity simulation (walk, stairs, sit, stand, lie) and a live prediction card.
• [Signal pattern gallery] Characteristic waveform of each of the six activities with plain-language notes on what distinguishes them.
• [Per-class evaluation view] Precision, recall and F1 per activity plus a confusion-matrix heatmap; numbers are design targets, computed for real by the notebook.
• [Windowing visualizer] Shows the 2.56-second sliding window with 50% overlap over the raw 50 Hz signal, the exact preprocessing the dataset uses.
• [Classical baseline comparison] An SVM on the 561 engineered features (the dataset paper's own baseline) alongside the LSTM, for the report's comparison section.
• [Error analysis section] Documents the sitting-vs-standing confusion and the effect of phone placement, straight from the held-out evaluation.

• Complete source code (signal pipeline, LSTM model, SVM baseline, windowing, demo console)
• Jupyter training and evaluation notebook (buyer-run procedure: load UCI HAR, train LSTM, evaluate on held-out subjects, compare with SVM)
• Project report PDF (background, signal preprocessing, architecture, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (why subject-disjoint splits, LSTM vs engineered features, sitting-vs-standing confusion, placement sensitivity)
• Setup guide (environment, dataset download, training your own model, running the console)