Human Activity Classification using Body Sensor Data

🎯 Project Highlights

• 94% Classification Accuracy across seven distinct human activities
• 12% Accuracy Improvement through advanced preprocessing and optimization
• Comprehensive ML Pipeline implementation for time-series data
• Robust Feature Engineering with multiple selection techniques

Project Overview

This project implements an advanced machine learning pipeline for human activity classification using multi-dimensional time-series data from body-worn sensors. The system processes and analyzes data from the AReM (Activity Recognition system based on Multisensor data fusion) dataset to classify seven distinct human activities.

📊 Technical Architecture

Data Engineering Pipeline

graph LR
    A[Raw Sensor Data] --> B[Preprocessing]
    B --> C[Feature Engineering]
    C --> D[Feature Selection]
    D --> E[Model Training]
    E --> F[Evaluation]

Loading

Advanced Feature Engineering

1. Time Series Processing

   • Dynamic segmentation up to 20 segments per series
   • Temporal pattern extraction
   • Local characteristic analysis

2. Statistical Feature Extraction

   features = {
       'statistical': ['min', 'max', 'mean', 'median'],
       'distribution': ['std', 'q1', 'q3'],
       'temporal': ['sliding_windows', 'segment_analysis']
   }

3. Dimensionality Reduction

   • Principal Component Analysis (PCA)
   • Recursive Feature Elimination (RFE)
   • Cross-validated feature selection

🛠 Technical Implementation

Class Imbalance Solutions

• SMOTE Implementation
  • Synthetic sample generation
  • Balanced class distribution
  • Enhanced model robustness

Model Optimization

1. Regularization Techniques

   • L1 (Lasso) for feature sparsity
   • L2 (Ridge) for overfitting prevention
   • Cross-validation for hyperparameter tuning

2. Performance Metrics

   metrics = {
       'accuracy': 0.94,
       'improvement': '12%',
       'cross_validation': '5-fold',
       'evaluation': ['confusion_matrix', 'ROC_curves']
   }

📈 Results and Impact

Classification Performance

• Overall Accuracy: 94%
• Improvement: 12% increase from baseline
• Cross-Validation: Consistent performance across folds

Feature Importance Analysis

key_features = {
    'temporal': ['segment_patterns', 'window_statistics'],
    'statistical': ['quartile_ranges', 'distribution_metrics'],
    'engineered': ['custom_indicators', 'derived_features']
}

Data Processing Pipeline

1. Raw Data Handling

   • 88 instances × 6 time series
   • 480 consecutive values per series
   • 7 distinct activity classes

2. Feature Engineering Framework

   • Statistical feature extraction
   • Temporal pattern analysis
   • Dimensionality optimization

3. Model Architecture

   • Regularized logistic regression
   • SMOTE for class balancing
   • Cross-validated parameter tuning

---