基于动态特征融合与熵感知采样的无人机目标检测方法OA

To address the challenges of varying object scales,severe occlusion,and complex lighting conditions in UAV aerial scenes,this paper proposes a multi-modal dynamic fusion-based object detection algorithm named DMSF-DETR.The method introduces innovative improvements to the Deformable-DETR framework:First,a Cross-scale Feature Fusion Module(CFFM)is designed to achieve efficient multi-scale feature fusion through bidirectional interaction paths and dynamic weight optimization mechanisms.Second,an entropy-aware dynamic sparse sampling attention mechanism is proposed to adaptively adjust feature sampling density,thereby improving the detection accuracy of small objects.Finally,a Dynamic Query Initialization(DQI)strategy is introduced,which leverages heatmaps to dynamically generate queries to enhance the recognition capability for occluded objects.Experimental results on the DroneVehicle dataset demonstrate that the proposed algorithm achieves a 0.94%improvement in mAP compared to the baseline model,exhibiting stronger robustness in complex scenarios such as nighttime and occlusion.This study provides an effective solution for multi-source information fusion-based object detection from a UAV perspective.