An illumination-aware visible-infrared fusion framework for rotated object detection on airborne platforms
Siyu Cheng, S Liu, Yuqi Han · IET conference proceedings. · 2026
Detecting rotating-bounded objects via visible-infrared (IR) fusion is vital for aerial vehicle detection in smart cities and disaster relief. Visible images often fail under extreme conditions like darkness or fog, while IR sensors offer low-light resilience. This study proposes an airborne platform algorithm: Firstly, the Illumination-Aware Cross-Modal Detection framework (IA-CMDet) employs a ResNeXt-FPN backbone with Convolutional Block Attention Module (CBAM) to enhance feature extraction robustness in complex environments. Secondly, the Oriented Region-based CNN (ORCNN) detector addresses arbitrary orientation challenges through oriented region proposal and refinement heads. A feature-level fusion method is also used to fuse information from two modalities at the feature level. Finally, Illumination-Aware Non-Maximum Suppression (IA-NMS) intelligently fuses outputs from independent visible, IR, and fused detection branches to resolve cross-modal information imbalance. Evaluations on the DroneVehicle dataset confirm significant detection rate improvements across challenging scenarios including dark nights and foggy conditions, demonstrating superior performance over conventional methods. Specifically, compared with the mAP of the baseline model Halfway Fusion, the mAP of our IA-CMDet improved by 6.91%.