道路交叉口自动检测算法的研究  

Research on automatic recognition algorithm of road intersection

作　　者：王龙飞 刘智[1] 金飞[1] 王番[1] WANG Longfei;LIU Zhi;JIN Fei;WANG Fan(Information Engineering University,Zhengzhou 450001,China;Troops 69089,Kuerle,Xinjiang 841000,China)

机构地区：[1]信息工程大学,郑州450001 [2]69089部队,新疆库尔勒841000

出　　处：《测绘科学》2020年第5期126-131,146,共7页Science of Surveying and Mapping

摘　　要：针对高分辨率遥感影像道路交叉口特征不明显、检测难度大等问题,该文提出一种改进的道路交叉口自动检测算法。该算法在YOLOv3网络基础上,首先采用参数修正单元激活卷积层,使目标特征在传递过程中保留更多负信息;然后在特征提取端与特征检测端之间实现多尺度特征融合,增强目标细节特征的提取;最后将单向卷积模块改进为多通道卷积模块,对卷积模块横向拉伸后再纵向聚合。为了验证算法有效性,对常见7种类型交叉口进行测试,实验结果表明:改进后算法对复杂背景下小尺寸道路交叉口的检测效果得到明显提升,有效实现了多种类型的道路交叉口自动化检测。Aiming at the problems of high-resolution remote sensing image road intersections with insignificant features and difficult detection,this paper proposes an automatic road intersection detection algorithm. This algorithm is based on the YOLOv3 network. First,the parameter correction unit is used to activate the convolution layer,so that the target features retain more negative information during the transfer process. Then,multi-scale feature fusion is realized between the feature extraction end and the feature detection end to enhance the extraction of the target detail feature. Finally,the unidirectional convolution module is improved to a multi-channel convolution module,and the convolution module is laterally stretched and then longitudinally aggregated. In order to verify the effectiveness of the algorithm,seven common types of intersections are tested. The experimental results show that the improved algorithm can significantly improve the detection of small-sized road intersections under complex background,and effectively achieved the automatic detection of various types of road intersections.

关 键 词：道路交叉口 YOLOv3网络 参数修正单元 多尺度融合 多通道卷积模块 

分 类 号：P237[天文地球—摄影测量与遥感]