RGB–infrared (IR) fusion is an effective way to improve object detection robustness for automotive perception under low-light and adverse-weather conditions. Yet, practical multi-modal detectors still face three issues: imperfect cross-modal alignment, inefficient long-range interaction, and unstable query initialization when modalities exhibit inconsistent evidence. This paper presents CMAFNet, a deployment-oriented cross-modal alignment and fusion network with three key designs. (1) A Dynamic Receptive Backbone (DRB) extracts multi-scale features with adaptive receptive fields for both modalities. (2) A Channel-Split Mamba Block (CSM-Block) models long-range cross-modal dependencies using selective state-space modeling with linear complexity in token length, enabling an efficient accuracy–latency trade-off. (3) A Global Multi-modal Interaction Network (GMIN) performs fine-grained alignment and adaptive fusion via dual-branch cross-attention guided by global average/max pooling. In addition, an uncertainty-minimal query selection strategy and a separable dynamic decoder further enhance detection stability and efficiency. Experiments on M3FD and FLIR-Aligned show that CMAFNet achieves 83.9% mAP50 and 84.2% mAP50, respectively, while maintaining competitive inference efficiency, supporting real-time automotive deployment on compute-constrained platforms.

Small underwater objects are often distinguished by weak contours and textures that multiscale aggregation can suppress. We propose DCP-UDet to preserve these local features as contextual information passes through the detector. Spatially embedded cross-channel attention extracts shared context, and wavelet–context refinement modulates local responses using directional detail. Gated scale fusion then controls the coarse features added to a fine-resolution residual path. During training, normalized scale-and-visibility weighting emphasizes difficult targets, while paired-view consistency regularizes predictions under photometric degradation. On the study-specific DUO split, DCP-UDet achieves 65.2% AP and 44.6% small-object AP, exceeding DEIM-D-FINE-S by 1.4 and 2.4 percentage points, respectively. The model has 14.8 million parameters and requires 36.2 GFLOPs, with a batch-one latency of 8.4 ms on an RTX A6000. Component and objective ablations show gains from both selective propagation and difficulty-normalized supervision. Under severe image corruption, the detector retains 79.8% of its clean-image AP. The results support refining coarse context before selectively adding it to fine features for underwater detection.