关键词: 无人机目标检测, 遮挡伪装识别, 多模态检测, 注意力机制, 对比学习

Abstract: To address the problems of severe occlusion, strong camouflage interference, and insufficient local valid evidence in visible-thermal object detection from unmanned aerial vehicles in mountainous forest environments, a dual-modal object detection method for complex scenes is investigated. An occlusion-camouflage collaborative perception dual-stream detection network is constructed by introducing an occlusion structure-aware visibility estimation mechanism, an occlusion-guided attention-enhanced fusion mechanism, and a local visible-part prototype and hard-background contrastive learning mechanism into a dual-stream detection framework. By jointly modeling the visibility, structural integrity, and boundary uncertainty of target regions, reliable regions in dual-modal features are explicitly selected. Through cross-modal attention interaction guided by joint visibility, together with channel and spatial recalibration, the residual target evidence is discriminatively enhanced. In addition, a local visible-part prototype and hard-background contrastive constraint is introduced during training to strengthen the feature discrimination between local target evidence and similar background regions, thereby improving object detection performance in complex scenes. Experimental results on the ODinMJ dataset show that the proposed method achieves detection accuracies of 65.8% and 68.6% under occlusion and camouflage conditions, respectively, and attains an overall average detection accuracy of 67.0% across all scenarios, exceeding the best competing method by 7.4%. On the DVTOD dataset, the proposed method achieves an average detection accuracy of 86.3%, outperforming the best competing method by 1.8%. While maintaining relatively low model complexity, the proposed method achieves superior detection performance and effectively improves the robustness and environmental adaptability of visible-thermal object detection from unmanned aerial vehicles in mountainous forest environments.

Key words: UAV object detection, occluded and camouflaged object detection, multimodal detection, attention mechanism, contrastive learning