Leveraging LDA feature extraction to augment human activity recognition accuracy

Human Activity Recognition (HAR) is a multiresearch-discipline area that integrates multiple-sensor data in recognizing and classifying all activities conducted physically by humans. These sensors can be part of the environment, wearable technology, or smartphones. The current research addresses the issues of reducing the dimensionality of the feature vector and accurate classification concerning smartphone-based human activity by proposing a hybrid feature extraction strategy that combines the LDA and MLP methods. Furthermore, SVM optimization with SGD has been used to increase the accuracy in activity classification. In this paper, LDA has been used to extract a new feature space that better enhances class separation and tests feature label prediction. The proposed approach, named LMSS, is tested against the UCI-HAR dataset and achieves state-of-the-art level accuracy. The proposed LDA-MLP-SVM-SGD (LMSS) framework achieves competitive state-of-the-art performance, reaching 99.52 % accuracy on the UCI-HAR dataset. Statistical analysis confirms that the difference versus the strongest baseline (mGRRF+XGB, 99.36 %) was not statistically significant (p = 0.12, α = 0.05), indicating comparable effectiveness rather than a definitive performance gap.