Human Activity Recognition (HAR) Project

Abstract

Human Activity Recognition (HAR) is a critical area of research with applications in health monitoring, sports analytics, and human-computer interaction. This project examines the performance of Naive Bayes and Support Vector Machine (SVM) classifiers in recognizing daily activities using accelerometer data collected from participants. Data preprocessing included feature scaling and addressing class imbalance using SMOTE. The analysis provides insights into the effectiveness of these algorithms in HAR systems.

Introduction

Human Activity Recognition (HAR) is essential in developing intelligent systems for monitoring and analyzing human behavior. This project focuses on using Naive Bayes and Support Vector Machine (SVM) to classify activities based on accelerometer data. The primary goal is to evaluate the accuracy and reliability of these algorithms in identifying various activities.

Dataset Description

Origin: The Human Activity Recognition Trondheim (HARTH) dataset.

Participants: 22 subjects wearing two Axivity AX3 accelerometers—one on the right front thigh and the other on the lower back.

Activities: Walking, running, sitting, standing, cycling, etc.

Attributes: Acceleration data in x, y, z directions for thigh and back sensors.

Methodology

Feature Engineering and Selection:

Data Cleaning: Handled missing values and ensured appropriate data types.

Normalization: Normalized accelerometer data to standardize the range of independent variables.

Feature Selection: Utilized six axes of accelerometer data along with timestamps and activity labels.

Model Training and Evaluation:

Naive Bayes:

Implemented using the Gaussian Naive Bayes classifier.

Trained on the preprocessed accelerometer data.

Support Vector Machine (SVM):

Implemented using a linear kernel.

Trained with feature scaling and hyperparameter tuning to optimize performance.

Evaluation Metrics: Accuracy, ROC AUC Score, Precision, Recall, F1-Score, Confusion Matrix.

Results and Analysis

Naive Bayes:

Accuracy: 68.9%

ROC AUC: 0.91

Analysis: Naive Bayes showed competitive performance with a ROC AUC score of 0.91, indicating good model discrimination. However, the accuracy of 68.9% suggests that while the model can differentiate between classes, it struggles with correct classifications for some activities. The simplicity of Naive Bayes makes it efficient, but it may not capture complex patterns in the data.

Support Vector Machine (SVM):

Accuracy: 46%

ROC AUC: 0.84

Analysis: The Naive Bayes achieved a higher accuracy of 68.9%, outperforming SVM in overall classification. Naive Bayes is effective in handling the multidimensional feature space of accelerometer data. However,Naive Bayes and SVM can be computationally intensive and may require significant tuning of hyperparameters to achieve optimal performance.

Discussion

The comparative analysis of Naive Bayes and SVM highlights the trade-offs between model simplicity and complexity. Naive Bayes, with its competitive ROC AUC, is a viable option for applications requiring fast, interpretable models despite its lower accuracy. SVM, while offering lower accuracy, demands more computational resources and fine-tuning. These findings suggest that the choice of algorithm depends on the specific requirements of the HAR application, such as the need for real-time processing versus the importance of higher accuracy.

Conclusion

This HAR project demonstrates the effectiveness of Naive Bayes and SVM classifiers in recognizing daily activities from accelerometer data. Naive Bayes offers a balance between simplicity and performance, Future work should explore hybrid models and advanced feature engineering to further improve HAR systems.