A Residual Alignment and Dynamic Multi-Kernel Network for Infrared–Visible Pedestrian Detection

To tackle feature redundancy, cross-modality misalignment, and scale sensitivity inherent in infrared–visible fusion-based pedestrian detection, this paper proposes a novel multimodal pedestrian detection network, termed EFDNet, with residual alignment and dynamic multi-kernel designs. Specifically, an enhanced multi-scale feature extraction (EMFE) module is developed to boost the discriminative capacity of fine-scale features. By embedding the efficient multi-scale attention (EMA) mechanism into a faster block architecture, this module strengthens edge and texture feature representation with marginal computational overhead. On this basis, a feature alignment restoration module (FARM) is constructed to suppress edge blurring and feature deviation caused by thermal radiation diffusion. Adopting a parallel residual-alignment-fusion paradigm with cross-modal dual-dimensional interaction, it achieves joint pixel-level and feature-level alignment across spatial and channel dimensions. Finally, a dynamic multi-kernel channel attention (DMCA) module is introduced for hierarchical feature refinement. Equipped with parallel multi-scale convolutions and a lightweight weight generation branch, this module optimizes feature characteristics from scale and channel perspectives, thus improving detection robustness for pedestrian targets of diverse scales. Extensive experiments on the KAIST dataset demonstrate that EFDNet achieves competitive L-AMR scores of 6.73%, 8.03%, and 4.11% on the overall, daytime, and nighttime subsets, respectively. Further evaluations on the FLIR and LLVIP datasets reveal that the proposed EFDNet attains mAP50, mAP75, and mAP50–95 of 82.8%, 36.6%, 42.0% and 97.3%, 79.0%, 68.2%, respectively, outperforming state-of-the-art methods on partial metrics and exhibiting competitive advantages.