融合GPNet与图像多尺度特性的红外-可见光道路目标检测优化方法  

Optimization of infrared-visible road target detection by fusing GPNet and image multiscale features

作　　者：孙文财[1] 胡旭歌 杨志发[1,2,3] 孟繁雨 孙微 SUN Wen-cai;HU Xu-ge;YANG Zhi-fa;MENG Fan-yu;SUN Wei(College of Transportation,Jilin University,Changchun 130022,China;Product Planning and Project Management Department,China FAW Group Co.,Ltd.,Changchun 130013,China;Changchun Automobile Industry Institute,Changchun 130031,China)

机构地区：[1]吉林大学交通学院,长春130022 [2]中国第一汽车集团有限公司产品策划及项目管理部,长春130013 [3]长春汽车工业高等专科学校,长春130031

出　　处：《吉林大学学报（工学版）》2024年第10期2799-2806,共8页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(51978310);吉林省交通运输创新发展支撑项目(2020-1-12)。

摘　　要：为提高道路交通安全领域中的道路目标检测的精度,本文借鉴图像融合技术中多尺度特征图像融合思想,融合和GPNet中Ghost瓶颈模块,实现了融合质量和较小算法复杂度的平衡,建立了一种红外和可见光融合及目标检测网络。网络分为选择性图像融合模块、轻量化目标检测模块和融合质量及检测精度判别网络3个部分。在白天、夜间和特殊天气(雨、雾等)下,平均车速30~40 km/h的城市工况下进行3组试验作为数据集,实验结果表明:平均梯度最高提升5.64881、交叉熵提升了0.93668、边缘强度提升了56.9457、信息熵提升了0.925208781、互信息提升了1.000548571、峰值信噪比提升了3.053893252、Qab提升了0.342882208、Qcb提升了0.20898381以及均方误差降低0.08。轻量化目标检测网络输出的AP、mAP和Recall均为最优水平,验证了红外和可见光技术应用在道路障碍物检测方面的优势。In order to improve the accuracy of road target detection in the field of road traffic safety,an innovative infrared and visible fusion and detection network is established by borrowing the idea of multi-level feature image fusion for fusion in image fusion technology and the idea of Ghost bottleneck module building in GPNet to reduce the complexity of the algorithm.The network is divided into three parts:selective image fusion module,lightweight target detection module and fusion quality and detection accuracy discriminative network.Three sets of experiments were conducted as data sets under urban working conditions with an average vehicle speed of 30-40 km/h in daytime,nighttime and special weather(rain,fog,etc.).Experimental results:the highest average gradient lift of 5.64881,cross-entropy of 0.93668,edge strength of 56.9457,information entropy of 0.925208781,mutual information of 1.000548571,peak signal-to-noise ratio 3.053893252,Qab0.342882208,Qcb0.20898381 and mean square error reduction of 0.08.The AP,mAP and Recall of the output of the lightweight target detection network are all at the optimal level,which verifies the advantages of the innovative application of infrared and visible light technologies for road obstacle detection.

关 键 词：交通运输系统工程 计算机视觉 红外和可见光图像融合 多尺度图像融合 YOLOv5目标检测 

分 类 号：U492.8[交通运输工程—交通运输规划与管理]