Generalization experiments on VisDrone datasets.

In the process of UAV small target vehicle detection, it is difficult to extract the features because of the small target shape of the vehicle, the environment noise is big, the vehicles are dense and easy to miss detection. The LMAD-YOLO model is proposed, and the MultiEdgeEnhancer module is designed to enhance the edge information and enhance the feature capture through a series of operations. Large Separable Kernel Attention and SPPF are combined to form MSPF module, which can realize multi-scale perception aggregation and improve the ability of distinguishing small targets from interference. Adown module is introduced to replace the model of sampling, in order to reduce the parameters and computational complexity while enhancing the accuracy of small target detection. A Multidimensional Diffusion Fusion Pyramid Network is designed, in which Dasi and feature spread mechanism are used to fuse features to reduce the error detection and missed detection. Compared with YOLO11n model P, R, MAP50 of the improved model on DroneVehicle data set were increased by 2.4%,1.4%,2.2% respectively. The model also showed good generalization ability on the VisDrone data set.在无人机小型目标车辆检测任务中，由于待检测车辆目标体型小巧、环境噪声干扰显著、车辆排布密集，不仅特征提取难度较高，还极易出现漏检问题。为此本文提出LMAD-YOLO模型，并设计了多边缘增强模块（MultiEdgeEnhancer），通过一系列操作强化边缘信息与特征捕获能力。将大型可分离卷积核注意力机制与SPPF模块相结合，构建得到MSPF模块，可实现多尺度感知聚合，提升区分小型目标与干扰信息的能力。引入Adown模块替代原有下采样模块，在提升小型目标检测精度的同时，降低了模型参数量与计算复杂度。此外，本文设计了多维扩散融合金字塔网络（Multidimensional Diffusion Fusion Pyramid Network），采用Dasi与特征扩散机制完成特征融合，以降低误检与漏检率。相较于YOLO11n模型，改进后的模型在DroneVehicle数据集上的精确率（P）、召回率（R）与MAP50分别提升2.4%、1.4%与2.2%；该模型在VisDrone数据集上同样展现出优异的泛化能力。