Human Activity Recognition (HAR) Mini Project

Overview

Implementation of a Human Activity Recognition system using accelerometer and gyroscope data. The dataset used is the UCI-HAR dataset, where participants performed six activities. The goal is to analyze and classify activities based on sensor data.

Dataset

• UCI-HAR Dataset
• Raw accelerometer data from the inertial_signals folder used.
• Data organized and sorted using CombineScript.py and MakeDataset.py scripts.

Preprocessing

1. Place CombineScript.py and MakeDataset.py in the UCI dataset folder.
2. Run CombineScript.py to organize data into the Combined folder.
3. Run MakeDataset.py to create a dataset with train, test, and validation sets.
4. Focus on the initial 10 seconds of activity (500 data samples at 50Hz).

Tasks

1. Waveform Analysis

• Plot waveform for each activity class.
• Observe differences/similarities in a subplot with 6 columns.

2. Static vs. Dynamic Activities

• Analyze linear acceleration for each activity.
• Justify the need for a machine learning model to differentiate static and dynamic activities.

3. Decision Tree Training and Evaluation

• Train Decision Tree using the train set.
• Report accuracy and confusion matrix using the test set.
• Train Decision Tree with varying depths (2-8) and analyze accuracy changes.

4. Feature Engineering and Visualization

• Use PCA on Total Acceleration for dimensionality reduction.
• Apply TSFEL for feature extraction.
• Visualize different activity classes using scatter plots.

5. Decision Tree with Engineered Features

• Train Decision Tree using features from TSFEL.
• Report accuracy and confusion matrix using the test set.
• Compare Decision Tree accuracies with varying depths using raw data and engineered features.

6. Model Performance Analysis

• Identify participants/activities with poor model performance.
• Analyze reasons for performance issues.

7. Deployment

• Utilize Physics Toolbox Suite to collect smartphone sensor data.
• Trim data to 10 seconds, ensuring consistent phone position and alignment.
• Train on UCI dataset and test on collected data.
• Report accuracy and confusion matrix.

Results

Confusion Matrix

Scatter Plots

• PCA Scatter Plot: Limited separability.
• TSFEL + PCA Scatter Plot: Improved class separability.

Decision Tree Depth Impact

• Optimal depth crucial for balancing bias-variance tradeoff.

Model Deployment

• Successful classification in real-world scenarios.
• Identified special cases impacting model performance.

Future Improvements

• Explore ensemble methods for dynamic activities.
• Investigate advanced feature engineering techniques.
• Implement real-time monitoring for dynamic environments.

Skills Demonstrated

• Data preprocessing, feature engineering, and model evaluation.
• Decision Tree implementation and analysis.
• PCA application for dimensionality reduction.
• Real-world deployment considerations.