On the Optimum Linear Soft Fusion of Classifiers

Applied Sciences, Journal Year: 2025, Volume and Issue: 15(9), P. 5038 - 5038

Published: May 1, 2025

We present new analytical developments that contribute to a better understanding of the (soft) fusion classifiers. To this end, we propose an optimal linear combiner based on minimum mean-square-error class estimation approach. This solution allows us define post-fusion improvement factor relative best fused classifier. Key elements for are number classifiers, their pairwise correlations, imbalance between performances, and bias. Furthermore, consider exponential models class-conditional probability densities establish relationship classifier’s error mean square estimate. predict reduction in These theoretical findings contrasted biosignal application detection arousals during sleep from EEG signals. The results obtained reasonably consistent with conclusions.

Language: Английский

A Privacy-Preserving AIoT Framework for Fall Detection and Classification Using Hierarchical Learning With Multi-Level Feature Fusion

IEEE Internet of Things Journal, Journal Year: 2024, Volume and Issue: 11(15), P. 26531 - 26547

Published: Aug. 1, 2024

Privacy and false fall detection pose a significant challenge within the current camera-based human activity monitoring research. In response, we propose solution that leverages inherent relationship between binary classification through hierarchical learning for low classification. Our involves employing edge computing data preprocessing, with specific focus on extracting key points of body motion features. Subsequently, process based multi-stream takes place at cloud level. This approach prevents sending raw RGB videos to improve privacy. Moreover, our network incorporates multi-class using stream-to-stream skip connections multi-level feature fusion. We created Multiscale Convolutional Fusion Block (MSCFB) extraction fusion an inner block connection. Additionally, added auxiliary loss first stream regulate control detection. attained state-of-the-art performance in UP-Fall dataset, securing F1-Score 97.27%. Notably, there were no misclassifications regular activities sub-classes, leading perfect 100% this dataset. Furthermore, achieved PRECIS HAR dataset 98.34%.

Language: Английский

Multi-Level Feature Exploration Using LSTM-Based Variational Autoencoder Network for Fall Detection

Lecture notes in computer science, Journal Year: 2024, Volume and Issue: unknown, P. 399 - 414

Published: Dec. 3, 2024

Language: Английский