Human Activity Recognition using Smartphones 📱🤸

This project performs classification of human activities (e.g., walking, sitting, standing) based on smartphone sensor data using machine learning models. The dataset consists of preprocessed accelerometer and gyroscope readings from 30 participants.

📦 Dataset

• Source: UCI HAR Dataset
• Samples: 10,299
• Features: 561 (time & frequency domain)
• Activities:
  • WALKING
  • WALKING_UPSTAIRS
  • WALKING_DOWNSTAIRS
  • SITTING
  • STANDING
  • LAYING

🔍 Key Phases

Data Understanding & EDA

• Class and subject distribution
• Feature correlation, PCA visualization

Preprocessing

• Label mapping
• Feature scaling (StandardScaler)
• SelectKBest feature selection
• PCA for dimensionality reduction (optional)

Modeling

• Trained: RandomForest, SVM (RBF), KNN, MLP
• Evaluated using accuracy and classification reports

Hyperparameter Tuning

• RandomizedSearchCV with StratifiedKFold for:
  • RandomForest
  • MLPClassifier
  • SVM (RBF)

Ensemble Techniques

• Soft Voting
• Stacking (meta-model: LogisticRegression)
• Blending (weighted average)

Error Analysis

• Confusion matrix
• Misclassified samples
• Most confused activity pairs

Final Output

• Accuracy: ~95–97% (ensemble models)
• Inference-ready prediction function using saved model, scaler, and selector