WCANet: Wavelet-Based Cross-Attention Framework for Multispectral Pedestrian Detection in Intelligent Transportation Systems

Fast, accurate, and robust pedestrian detection is crucial for intelligent transportation systems. To address the limitations of single-modality detectors in complex traffic scenarios, we propose WCANet, a novel multispectral pedestrian detection framework integrating visible and infrared images for all-weather monitoring. Specifically, we design a frequency-domain interaction and spatial-domain fusion framework to integrate multi-level visible and infrared features extracted by a dual-stream network. First, we propose the Wavelet-based Cross-Attention (WCA) module, which employs discrete wavelet transform (DWT) to decouple low-frequency homogeneous feature interaction and high-frequency complementary feature preservation. We develop the Iterative High-Low Cross-Attention (I-HLCA) module, which innovatively introduces high-frequency attention into the low-frequency feature interaction process. Additionally, multi-head dilated convolutions are used to explore local spatial similarities in cross-modality high-frequency features. Furthermore, we design a channel attention-guided visible-infrared spatial feature fusion module, enabling joint optimization across frequency and spatial domains to comprehensively enhance multispectral feature representation. Comprehensive experiments demonstrate that WCANet achieves state-of-the-art performance on both KAIST and LLVIP multispectral pedestrian detection benchmarks, delivering stable and reliable detection.