DS⁃YOLO:一种部署在无人机终端上的小目标实时检测算法

DS⁃YOLO:a real⁃time small object detection algorithm on UAVs

随着无人机硬件成本的降低和深度学习算法的发展,部署在无人机终端的实时目标检测算法在诸多领域展现出广泛的应用前景。然而,嵌入式设备有限的能耗和算力,以及普适性目标检测算法对于小目标特征提取不够充分等问题,制约了此类算法速度和精度的提升。文中提出了一种部署在无人机终端上的小目标实时检测算法DS⁃YOLO(Dense⁃SPP YOLO),算法基于密集连接的思想设计了全新的主干网络,并改进了空间金字塔池化模块以增强小目标的特征提取和多尺度特征复用,最后基于批归一化层(Batch⁃Normalization)的缩放因子修剪网络中不重要的通道,修剪瘦身后的算法更加适合部署在移动端。在Visdrone2019⁃DET数据集上的测试结果表明,DS⁃YOLO算法mAP(mean Average Precision)指标比SlimYOLOv3算法提升约3%,检测速度达到89 FPS(Frames Per Second),高于SlimYOLOv3的67 FPS。

Real⁃time object detection algorithms deployed on unmanned aerial vehicles(UAVs)have the broad application prospects in many fields with the reduction of UAVs hardware costs and the development of deep learning algorithms.However,the limited energy consumption and the computing ca⁃pability of embedded devices,as well as the inadequate feature extraction of small target features by com⁃mon object detection algorithms,constrain the enhancements of accuracy and speed of the algorithms.A novel real⁃time small object detection algorithm on UAVs,Dense⁃SPP⁃YOLO(DS⁃YOLO),is proposed The algorithm adopts a new backbone network based on dense connections,and designs a modified spatial pyramid pooling(SPP)module to enhance the small object feature extraction and the multi⁃scale feature fusion.Then,the unimportant channels in the network are pruned based on the scaling factor of each batch⁃normalization layer,making the algorithm thinner to adapt to the deployment on mobile terminal.The DS⁃YOLO on Visdrone2019⁃DET dataset is evaluated.Detection results demonstrate that compared with the SlimYOLOv3 algorithm,the mean average precision(mAP)of the DS⁃YOLO algorithm increases about 3%and the detection speed reaches 89 FPS(frames per second),it is high than that of the SlimYOLOv3 algorithm(67 FPS).