Hidden Markov Model for Human Activity Recognition

Project Overview

This project implements a Hidden Markov Model (HMM) to classify human activities using smartphone sensor data from accelerometer and gyroscope. The model distinguishes between motion patterns with 84.90% training accuracy and 65.79% test accuracy on unseen data.

---

Group Members

• Excel – Data Collection, Feature Engineering, HMM Implementation
• Lesly – Data Collection, Model Evaluation, Report Writing

---

Project Structure

HMM-Activity-Recognition/
├── data/                          # Raw sensor data (258 files)
│   ├── jumping/                   # 10 sessions
│   ├── standing/                  # 10 sessions
│   ├── still/                     # 10 sessions
│   ├── walking/                   # 12 sessions
│   └── test/                      # Unseen test data
├── notebook/
│   └── hmm_activity_recognition_final.ipynb
├── models/                        # Saved models and scalers
│   ├── hmm_model_20251031_180138.pkl
│   ├── scaler_20251031_180138.pkl
│   ├── feature_names_20251031_180138.json
│   └── training_metrics_20251031_180138.json
├── report/
│   └── report.pdf/                   
├── metrics       
└── plots        

---

Key Results

• Training Accuracy: 84.90% (all 4 activities)
• Test Accuracy: 65.79% (on Jumping & Walking)
• Best Performing Activity: Walking (81.58% sensitivity)
• Activities Recognized: Jumping, Standing, Still, Walking
• Model Saved: Complete HMM with feature pipeline

---

Performance Summary

Training Performance

Activity	Samples	Sensitivity	Specificity
Jumping	406	50.99%	99.92%
Standing	402	95.52%	97.25%
Still	360	97.50%	100.00%
Walking	508	94.69%	81.42%

Test Performance (Unseen Data)

Activity	Samples	Sensitivity	Specificity
Jumping	38	70.00%	100.00%
Walking	38	81.58%	50.00%

---

Model Architecture

• Type: Hidden Markov Model (HMM) with Gaussian emissions
• States: 4 (Jumping, Standing, Still, Walking)
• Features: 23 per analysis window
• Training Algorithm: Baum-Welch
• Decoding Algorithm: Viterbi

---

Features Extracted

• Time-domain: Mean, Standard Deviation, RMS, SMA, Correlations, Magnitude
• Frequency-domain: Dominant Frequency, Spectral Energy, Spectral Entropy

---

Technical Implementation

• Window Size: 100 samples (1 second at 100 Hz)
• Overlap: 50%
• Normalization: Z-score standardization
• Training Samples: 1,676 feature vectors
• Test Samples: 76 feature vectors

---

Usage

# Run the complete pipeline in Jupyter Notebook
jupyter notebook notebook/hmm_activity_recognition_final.ipynb

---

Requirements

numpy, pandas, matplotlib, seaborn, scipy, scikit-learn

---

Model Files Generated

• hmm_model_20251031_180138.pkl – Trained HMM parameters
• scaler_20251031_180138.pkl – Feature normalization scaler
• feature_names_20251031_180138.json – Feature names
• training_metrics_20251031_180138.json – Training statistics
• evaluation_metrics.csv – Performance metrics