Optimized human detection in NLOS scenarios using hybrid dimensionality reduction and SVM with UWB signals

Abstract

Trapped victim localization in search and rescue (SAR) operations is especially difficult in non-line-of-sight (NLOS) conditions, where traditional techniques fail due to debris and signal distortion. Ultra-wideband (UWB) NLOS signal datasets offer a promising alternative but are often high-dimensional and noisy. This study proposes an optimized dimensionality reduction framework combining an adaptive human presence detector (AHPD) with genetic algorithms (GA) and independent component analysis (ICA), followed by support vector machine (SVM) classification. The approach is tested on a public NLOS dataset comprising 23,522 dynamic instances, each with 256 signal samples per attribute, simulating complex SAR scenarios including rubble and dynamic obstacles. The results indicate that the AHPD+GA+SVM model reached an accuracy of 85.78%, sensitivity of 80.00%, and specificity of 96.46%, which is better than the AHPD+ICA +SVM model that had an accuracy of 79.20%, sensitivity of 73.07%, and specificity of 81.05%. These findings demonstrate the framework’s robustness and scalability, making it a strong candidate for real-time human detection in disaster recovery missions.