Human Activity Recognition System (HARS)

This project focuses on Human Activity Recognition (HAR) using data collected from smartphones’ embedded sensors such as accelerometers and gyroscopes. The task is to classify physical activities (e.g., walking, sitting, standing, etc.) based on extracted features.

📌 Dataset Info

• Train set: 7,353 samples
• Test set: 1,000 samples
• Features: 562 sensor-derived features (accelerometer, gyroscope, etc.)
• Activity classes:
  • Laying
  • Standing
  • Sitting
  • Walking
  • Walking Upstairs
  • Walking Downstairs

🔬 Methodology

1. Data Preprocessing & Feature Engineering

   • Normalization and cleaning
   • Dimensionality reduction explored via PCA and t-SNE

2. Train/Test Split

   • Dataset split in 80:20 ratio

3. Models Trained & Tuned

   • Logistic Regression

     • Parameters: { 'max_iter': [100, 200, 500] }
     • Best: max_iter=500
     • Accuracy: 95.5%

   • Support Vector Machine (SVM)

     • Parameters: { 'kernel': ['linear', 'rbf', 'poly', 'sigmoid'], 'C': [100, 50] }
     • Best: kernel='rbf', C=50
     • Accuracy: 96.6%

   • Decision Tree

     • Parameters: { 'max_depth': [2, 4, 6, 8] }
     • Best: max_depth=6
     • Accuracy: 83.6%

   • Random Forest

     • Parameters: { 'n_estimators': [20–100], 'max_depth': [2–16] }
     • Best: n_estimators=100, max_depth=14
     • Accuracy: 91.6%

📊 Results

• Best performing model: SVM (RBF kernel, C=50) with 96.6% accuracy
• Logistic Regression also performed strongly with 95.5% accuracy.
• Ensemble methods (Random Forest) showed competitive results but slightly lower accuracy compared to SVM.
• Decision Tree underperformed relative to other approaches.

🛠 Technologies and Packages Used

• Python (Jupyter Notebook)
• NumPy, Pandas, Matplotlib, Seaborn
• scikit-learn (Logistic Regression, SVM, Decision Tree, Random Forest)

---

👤 Author

Developed by Dipean Dasgupta
BTech CSE Graduate, AI & ML Enthusiast