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
  
DOI:
中文关键词:  无人机;小目标检测;密集连接;通道剪枝
英文关键词:unmanned aerial vehicles(UAVs); small object detection; dense connection; channel pruning
基金项目:国家重点研发计划(2019YFB2101700)资助项目
作者单位
张伟 南京邮电大学 计算机学院,江苏南京210023 
庄幸涛 南京邮电大学 计算机学院,江苏南京210023 
王雪力 南京邮电大学 计算机学院,江苏南京210023 
陈云芳 南京邮电大学 计算机学院,江苏南京210023 
李延超 南京邮电大学 计算机学院,江苏南京210023 
摘要点击次数: 227
全文下载次数: 61
中文摘要:
      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89FPS(Frames Per Second),高于SlimYOLOv3的67FPS。
英文摘要:
      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 capability of embedded devices, as well as the inadequate feature extraction of small target features by common 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).
查看全文  查看/发表评论  下载PDF阅读器

你是第2143246访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计