Multi-scale contactless sensing : adaptive signal processing and feature fusion for robust recognition

Abstract

With the continuous advancement of artificial intelligence (AI) and the Internet of Things (IoT), contactless sensing has emerged as a pivotal technology, offering non-intrusive, high-resolution monitoring across diverse applications. However, ensuring robust real-time sensing in dynamic and complex environments remains a significant challenge. Current systems face three major bottlenecks. First, at the signal preprocessing stage, multipath interference and noise degrade the fidelity of weak signal components, undermining the quality of downstream analysis. Second, during feature extraction, signals originating from multiple spatial regions and hierarchical levels become intertwined with irrelevant information, making it challenging to isolate salient features critical for accurate recognition. Finally, in the feature fusion stage, conventional single-scale approaches struggle to integrate transient fluctuations with long-term trends, limiting the system's ability to perform effective multi-modal and cross-temporal data fusion. To address these challenges, this study develops a contactless sensing framework that integrates adaptive signal processing and feature fusion to enhance robustness, precision, and scalability. First, we propose an adaptive signal processing pipeline, incorporating a parametrically tunable windowing function and frequency-domain transformation strategy. This method dynamically balances spectral resolution and side-lobe suppression, ensuring high-fidelity signal acquisition even in low signal-to-noise ratio (SNR) conditions. For feature extraction, we introduce a hierarchical feature selection and refinement mechanism, which iteratively enhances target signals while mitigating background interference. Furthermore, a cross-domain feature fusion strategy is employed, leveraging generative projection and domain-adaptive reconstruction to reduce noise sensitivity while preserving discriminative features. Compared to conventional filtering or threshold-based methods, this approach eliminates anomalies dynamically without relying on static parameters or pre-defined priors, ensuring generalizability across diverse sensing conditions. To simultaneously capture short-term variations and long-term trends, we propose a multi-resolution temporal fusion strategy, aligning information across different time scales into a unified analytical model. This mitigates the limitations of single-scale processing, allowing the system to maintain robust recognition performance in dynamically evolving environments, particularly for physiological state monitoring. By integrating these methodologies, the proposed sensing system achieves real-time efficiency, high-resolution signal processing, and robust recognition under complex, multi-source interference conditions. Ultimately, this study establishes a scalable, adaptive, and high-precision contactless sensing paradigm, advancing its practical applicability in real-world intelligent sensing and monitoring systems.