Human Activity Recognition with Smartphones

Overview

This project utilizes machine learning to classify human activities based on smartphone sensor data. Using a dataset collected from 30 volunteers performing activities ranging from static postures to dynamic movements, we employ various algorithms to accurately identify the activity type.

Dataset

The dataset comprises 3-axial linear acceleration and 3-axial angular velocity data captured at a 50Hz sampling rate from a Samsung Galaxy S II smartphone worn by the subjects. Activities include six basic types, alongside postural transitions.

Technologies Used

Programming Languages: Python

Key Libraries:
- NumPy
- pandas
- Matplotlib
- seaborn
- scikit-learn
- statsmodels
- ydata_profiling
Machine Learning Models:
- Random Forest
- XGBoost
- Linear SVM
- KNN
- Voting Classifier

Loading the dataset

df_test = pd.read_csv('activity_test.csv') df_train = pd.read_csv('activity_train.csv')

Basic dataset information

df_train.info() Feature Engineering A comprehensive feature set is extracted using time and frequency domain variables, applying noise filters, and separating the acceleration signal into body and gravity components.

Model Training and Evaluation We employ RandomForestClassifier for the initial model training, followed by PCA for feature dimensionality reduction. Model performance is evaluated using accuracy, precision, and recall metrics, supplemented by confusion matrix visualization.

Analysis and Results The project includes an in-depth analysis of model performance, highlighting the importance of feature engineering and selection in improving classification accuracy.