1.厦门理工学院;2.厦门大学
福建省自然科学基金资助项目（2023J01439）
School of Mechanical and Automotive Engineering
upported by the Natural Science Foundation of Fujian Province（2023J01439）
针对航拍密集交通图像中存在的小尺度目标检测精度低和模型参数量大的问题，提出一种轻量、高效的航拍图像检测算法模型UAVDet。首先设计了大核可分离注意力空间池化模块LSKASPM来加强模型对小尺度目标空间信息和语义信息的捕捉能力；其次构建了可变形上下文特征引导聚合模块C2f-DCG来增强模型对各尺度目标的特征理解能力；然后提出多尺度特征融合模块MSFM为引入的高分辨率检测分支SHead聚和更具细粒度的全局特征；最后应用基于网络权重幅值的层自适应稀疏剪枝技术LAMP降低模型参数规模。在公开数据集VisDrone上的实验结果表明，该模型针对城市十类常见交通目标的平均检测精度和漏检率分别为47.2%和47.5%，模型参数量为6.3M，推理速度达到197帧/秒，均优于现有公开算法。相关算法代码将在https://github.com/XMUT-Vsion-Lab/UAVDet公开。
To address the issues of low detection accuracy for small-scale objects and large model parameters in aerial images of dense traffic, we propose a lightweight and efficient aerial image detection algorithm model, UAVDet. First, we design the large-kernel separable attention spatial pooling module (LSKASPM) to enhance the model's ability to capture spatial and semantic information for small-scale objects. Next, we construct the deformable context feature-guided aggregation module (C2f-DCG) to improve the model's feature understanding across multiple scales. Then, we introduce the multi-scale feature fusion module (MSFM) to aggregate high-resolution detection branch (SHead) features and provide more fine-grained global features. Finally, the layer-wise adaptive sparse pruning technique (LAMP) based on network weight magnitudes is applied to reduce the model's parameter size. Experimental results on the public VisDrone dataset show that the model achieves an average detection accuracy of 47.2% and a missed detection rate of 47.5% for ten common traffic target classes in urban areas. The model has 6.3M parameters and an inference speed of 197 frames per second, outperforming existing public algorithms. The relevant algorithm code will be publicly available at https://github.com/XMUT-Vsion-Lab/UAVDet.
