MDFDET: A Multi-Modal Dynamic Fusion Algorithm for RGB-Infrared Object Detection

In low-illumination traffic scenarios, multimodal object detection encounters challenges such as weak cross-modal interaction, insufficient detail and boundary fidelity in upsampling, and limited global selection with noise suppression, leading to suboptimal multisource complementarity. This paper proposes a multi-modal dynamic perception fusion detection algorithm based on visible and infrared modalities. A cross-layer RGB-Infrared mid-level fusion backbone is designed to mitigate single-modal degradation and domain shift, enabling deeper interaction. To enhance representation, a C3M (Cross-Stage Convolutional-Mixed Multi-scale Attention Module) addresses scale and semantic deficiencies, while DySample is embedded in the neck for adaptive upsampling, preserving detail fidelity. Furthermore, a GSCSA (Grouped Spatial-Channel Synergistic Attention) module is added before the detection head to suppress semantic mismatch and environmental noise. Extensive experiments on public traffic datasets validate the framework: it improves mAP50:95 by 8.3% on M3FD dataset and by 5.6% on DroneVehicle dataset compared with state-of-the-art methods, achieving superior accuracy without compromising efficiency.