Leveraging LDA feature extraction to augment human activity recognition accuracy.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for understanding human behavior using sensor data.
• Smartphone-based HAR faces challenges in feature dimensionality reduction and classification accuracy.
• Existing methods require robust feature extraction and optimized classification models.

Purpose of the Study:

• To propose a hybrid feature extraction strategy combining Linear Discriminant Analysis (LDA) and Multilayer Perceptron (MLP) for smartphone-based HAR.
• To enhance classification accuracy by optimizing Support Vector Machine (SVM) with Stochastic Gradient Descent (SGD).
• To validate the proposed LDA-MLP-SVM-SGD (LMSS) approach on the UCI-HAR dataset.

Main Methods:

• Feature extraction using a hybrid LDA and MLP approach to improve class separability.
• Classification using an SVM model optimized with SGD for enhanced accuracy.
• Implementation of the LMSS framework for human activity recognition.

Main Results:

• The proposed LMSS framework achieved a state-of-the-art accuracy of 99.52% on the UCI-HAR dataset.
• LDA effectively extracted a new feature space enhancing class separation and prediction.
• Statistical analysis showed the LMSS approach is comparable to the strongest baseline (mGRRF+XGB, 99.36%).

Conclusions:

• The LMSS framework demonstrates a highly effective hybrid strategy for smartphone-based HAR.
• The combination of LDA, MLP, and optimized SVM with SGD significantly improves activity classification accuracy.
• The approach offers competitive performance, achieving state-of-the-art results with statistical comparability to existing top methods.