隧道内不均匀照度下无人驾驶视觉融合感知方法  

Visual fusion perception method under uneven illumination in tunnel of autonomous vehicles

作　　者：马庆禄[1] 汪曦洪 马恋 段学锋 MA Qinglu;WANG Xihong;MA Lian;DUAN Xuefeng(School of Traffic&Transportation,Chongqing Jiaotong University,Chongqing 400074,China;Ningxia Jiaotou Expressway Management Co.,Ltd.,Yinchuan 401120,China)

机构地区：[1]重庆交通大学交通运输学院,重庆400074 [2]宁夏交投高速公路管理有限公司,宁夏银川401120

出　　处：《应用光学》2025年第1期89-101,共13页Journal of Applied Optics

基　　金：交通部三峡库区奉建高速公路安全智能建造科技示范工程(Z29210003);重庆交通大学研究生科研创新项目(CYS240483)。

摘　　要：针对隧道内不均匀光线引起的无人驾驶汽车视觉感知难题,提出采用主动式红外与可见光相融合的增强方法。通过补充图像纹理和细节信息,改善隧道内光线波动对无人驾驶的影响。首先,利用导向滤波对图像进行去噪处理,并通过对比度受限的自适应直方图均衡化对不同照度下的视觉信息进行增强;其次,对增强后的红外与可见光图像利用非下采样轮廓波方法进行分解,将区域能量自适应加权平均和卷积稀疏表示的融合规则分别应用于低频基础和细节部分,对高频细节分量应用显著性融合规则;最后进行重构得到融合图像。实验结果表明:基于卷积稀CSR-RE算法(convolutional sparse representation and region energy adaptive Weighting)对比DLF(deep learning framework)算法平均时间减少了0.05 s,相较于基于稀疏表示的融合方法MI(mutual information)值提升了1.5,SF(spatial frequency)值最高提升了0.51,整体性能优于其他算法。In order to solve the visual perception problem of unmanned vehicles caused by uneven light in tunnels,an enhancement method based on the fusion of active infrared and visible light was proposed.By supplementing image texture and detail information,the impact of tunnel light fluctuations on unmanned driving was improved.Firstly,the image was denoised by pilot filter,and the visual information in the tunnel under different illuminance was enhanced by contrast limit adaptive histogram equalization.Secondly,the enhanced infrared and visible images were decomposed by using the non-subsampling contour wave method.The fusion rules of regional energy adaptive weighted average and convolutional sparse representation were used for low frequency base and detail respectively,and the significance fusion rules were applied to the directional components of high frequency details.Finally,the high and low frequency fusion components were reconstructed.Experimental results show that the average time of CSR-RE algorithm based on convolutional sparse representation and regional energy adaptive weighted average is reduced by 0.05 s compared with deep learning framework(DLF)algorithm.Compared with the fusion method based on sparse representation,the mutual information(MI)value of CSR-RE algorithm is increased by 1.5,and the spatial frequency(SF)value is increased by 0.51 at most.The overall performance of CSR-RE algorithm is better than that of other algorithms.

关 键 词：图像融合 公路隧道 红外图像 卷积稀疏表示 

分 类 号：TN201[电子电信—物理电子学] TP391.41[自动化与计算机技术—计算机应用技术]