基于神经网络的无人机影像多目标检测方法  

Multi Target Detection Method of Drone Image

作　　者：杜祥宇 DU Xiangyu(Beijing Institute of Surveying and Mapping, Beijing 100038, China;Beijing Key Laboratory of Urban Spatial Information Engineering, Beijing 100038, China)

机构地区：[1]北京市测绘设计研究院,北京100038 [2]城市空间信息工程北京市重点实验室,北京100038

出　　处：《北京测绘》2021年第10期1283-1287,共5页Beijing Surveying and Mapping

基　　金：国家重点研发计划(2016YFF0201305,2016YFF0201301)。

摘　　要：针对现有方法检测无人机目标时精度与泛化能力不足,提出一种基于回归的多目标检测方法。使用密集连接增强层间信息传递,添加批量再规范化(Batch Renormalization,BRN)层加速模型训练,降低样本分布不均而导致的精度偏低,使用密集连接结构聚合上采样层不同层信息。以开源数据集Vision Drone为基础建立优化数据集训练模型。结果表明,提出模型检测精度达89.57%平均精度均值,相比(You Only Look Once v3,YOLOv3)模型和(Region Full Convolutional Network,R-FCN)分别提高6.53%和3.11%,检测速度达27每秒传输帧数,在不同场景表现稳定。Aiming at the insufficient accuracy and generalization ability of existing methods to detect UAV targets,a regression-based multi-target detection method was proposed.Dense connections to enhance information transfer between layers were applied,and a Batch Renormalization(BRN)layer was added to accelerate model training to reduce the low accuracy caused by uneven were explored sample distribution.The dense connections to aggregate information from different layers of the upsampling layer and an optimized data set training model based on the open source data set Vision Drone was established.The results showed that the proposed model had an average accuracy of 89.57%,which was 6.53%and 3.11%higher than the(You Only Look Once v3,YOLOv3)model and(Region Full Convolutional Network,R-FCN),and the detection speed was 27 transmission frames per second,stable performance in different scenes.

关 键 词：神经网络 无人机航拍影像 多目标检测 密集连接 

分 类 号：P237[天文地球—摄影测量与遥感]