Abstract: To tackle the challenges of weak sensing capacity and high missed detection rates for small-scale objects using low-altitude UAV in complex traffic scenarios, the LMUAV-YOLOv8 algorithm is proposed. Its efficiency and advantage are verified through ablation and comparative experiments. The internal mechanisms is visualized by using the method of class activation mapping. In this dissertation, a lightweight feature fusion network (UAV_RepGFPN) is introduced firstly, proposing new feature fusion paths and a feature fusion module DBB_GELAN, which reduces the number of parameters and computation while improving the performance of the feature fusion network. Secondly, the feature extraction module (FTA_C2f) is constructed using partial convolution (PConv) and triplet attention mechanism (Triplet Attention), and the ADown down-sampling module is introduced. By rearranging the dimensions of the input feature maps and making fine-grained adjustments, the ability of the deep network to capture spatial features is enhanced, further reducing the number of parameters and computation. Then, concerning large amount of information loss during in layer-by-layer feature extraction and spatial transformation, a new context-guided programmable gradient information (UAV_PGI) strategy is proposed. By designing a context-guided reversible architecture and three additional auxiliary detection heads, UAV_PGI significantly enhance detection capabilities for aerial objects. In order to verify the validity and generalization ability of the model, comparative experiments are carried out on the VisDrone 2019 test set, and the results show that: compared with YOLOv8s, LMUAV-YOLOv8s on the VisDrone 2019 test set improves precision, recall, mAP@0.5, and mAP@0.5:0.95 by 4.2, 3.9, 5.1, and 3.0?percentage points, separately, with the computational cost increased by only 0.4?GFLOPs and the parameter count reduced by 63.9%, meaning a good balance between performance and cost. The inference experimental results based on NVIDIA Jetson Xavier NX embedded platform show that compared with the baseline model, the proposed algorithm can obtain higher detection accuracy under the condition of meeting the requirements of real-time detection, rendering it more suitable for real-time target detection scenarios in drones. Finally, the decision making process is visualized by using the method of class activation mapping, which provides a intuitive way to understand the internal mechanisms of the networ. And the results show that the proposed model has superior small-scale feature extraction and high-resolution processing capabilities.

Key words: small object detection, multi-scale, lightweight, YOLOv8, programable gradient information

摘要： 针对低空无人机目标检测面临目标尺度变化大、小目标容易漏检和误检的挑战，发展了一种融合多尺度特征的目标检测轻量化网络（LMUAV-YOLOv8），通过开展消融和对比实验，验证了算法的有效性和先进性，并借助类激活图，对模型的决策过程进行了解释。设计了一种轻量化的特征融合网络（UAV_RepGFPN），提出新的特征融合路径以及特征融合模块DBB_GELAN，降低参数量和计算量的同时，提高特征融合网络的性能。使用部分卷积（PConv）和三重注意力机制（Triplet Attention）构建特征提取模块（FTA_C2f），并引入ADown下采样模块，通过对输入特征图维度的重新排列和细粒度调整，以提升模型中深层网络对空间特征的捕捉能力，并进一步降低参数量和计算量。优化YOLOv9的可编程梯度信息（programmable gradient information，PGI）策略，设计基于上下文引导（Context_guided）的可逆架构，并额外生成三个辅助检测头，提出UAV_PGI可编程梯度方法，避免传统深度监督中多路径特征集成可能导致的语义信息损失。为了验证模型的有效性及泛化能力，在VisDrone 2019测试集上开展了对比实验，结果显示，与YOLOv8s相比，LMUAV-YOLOv8s的准确度、召回率、mAP@0.5和mAP@0.5：0.95等指标分别提升了4.2、3.9、5.1和3.0个百分点，同时参数量减少了63.9%，计算量仅增加0.4?GFLOPs，实现了检测性能与资源消耗的良好平衡。基于NVIDIA Jetson Xavier NX嵌入式平台的推理实验结果显示：与基线模型相比，该算法能够在满足实时检测要求的条件下，获得更高的检测精度，对于无人机实时目标检测场景具有较好的适用性。借助类激活图，对算法的决策过程进行了可视化分析，结果表明，该模型具备更优异的小尺度特征提取和高分辨率处理能力。

关键词: 小目标检测, 多尺度, 轻量化, YOLOv8, 可编程梯度信息