FDDF: Frequency decomposition and spatial-frequency dual-domain fusion network for multi-spectral pedestrian detection
Xiaowei Liu, Gang Xie, Xinlin Xie, Xinying Xu · Expert Systems with Applications · 2026
Multispectral pedestrian detection plays a pivotal role in advancing the safety and reliability of autonomous driving and surveillance systems, particularly under low-light scenarios or adverse weather conditions. However, existing methods predominantly rely on exploiting spatial domain features from only two spectrums, which fails to explore the complementary relationship between spatial and frequency domains, limiting the capture of complex inter-modal information. To address this limitation, we establish a novel dual-domain paradigm for multispectral pedestrian detection, realized through the proposed Frequency Decomposition and Spatial-Frequency Dual-Domain Fusion Network (FDDF). This paradigm extends beyond conventional image-domain fusion by explicitly incorporating frequency-domain representations to complement spatial-domain features. FDDF introduces a learnable Gaussian-based Frequency-Domain Feature Decomposition module to adaptively separate high- and low-frequency responses, followed by a Frequency Spectrum Attention mechanism that emphasizes informative spectral components while suppressing modality-specific noise. Finally, Frequency-Spatial Domain Feature Global Co-occurrence(FSC) is introduced to realize the fusion and alignment of spatial domain and frequency domain features. Extensive experiments on KAIST, CVC-14, and LLVIP benchmarks demonstrate that FDDF achieves competitive or superior performance under both loose and strict IoU criteria. Our method provides a novel cross domain fusion solution for multispectral pedestrian detection.