高精度三维路面纹理超分辨率重构及测评方法
Super-resolution Reconstruction and Evaluation Method of High-precision Three-dimensional Pavement Texture
战友 1林修全 1邱延峻 1艾长发 1张傲南1
作者信息
- 1. 西南交通大学土木工程学院,四川 成都 610031;道路工程四川省重点实验室,四川 成都 610031
- 折叠
摘要
路面抗滑性能受路面宏观纹理和微观纹理共同影响,为了实现车载激光扫描设备快速收集高分辨率三维路面纹理,以及路面抗滑性能的连续非接触式微观尺度测评,构建了基于自监督深度学习的超分辨率网络模型,循环递归地将不同低分辨率的路面纹理沿行车方向重构至0.1 mm精度.将527幅0.1 mm分辨率的SMA-13沥青路面纹理用于网络训练与测试,利用最邻近法分别以1/2、1/4、1/8、1/16的下采样因子模拟不同车速下的低分辨率纹理.将分辨率相差2倍的同种纹理对作为网络模型的输入,并按8:2比例随机划分训练集和测试集,训练集通过小尺寸分割进行数据增强.在最优权重下,模型能够循环递归地将不同低分辨率的路面纹理重构至0.1 mm·像素-1分辨率.利用峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)、结构相似度(Structural Similari-ty,SSIM)和平均构造深度相对误差评估超分辨率纹理的重构质量,并以双三次插值法作为对照.结合双三次插值法与所提出的超分辨率网络,研究非预设下采样因子模型对低分辨率路面纹理的重构性能.最后,基于OOA-LightGBM算法,提取高分辨率纹理以及超分辨率重构纹理的特征参数,构建抗滑性能预测模型,研究基于超分辨率重构纹理预测抗滑性能的可能性.结果表明:所提出的超分辨率网络有助于车载激光扫描设备在128 km·h-1的行驶速度内采集PSNR大于30 dB、SSIM大于0.95、平均构造深度相对误差绝对值小于1％的0.1 mm高精度路面三维纹理;在16、32、64 km·h-1的恒定行驶速度下收集的路面三维纹理,其抗滑性能预测的平均决定系数R2为 0.808.
Abstract
The skid resistance of pavement surfaces is influenced by both macro-texture and micro-texture.To achieve high-resolution 3D pavement texture acquisition using vehicle-mounted laser scanning equipment and enable continuous,non-contact microscale evaluation of pavement skid resistance,this study constructed a super-resolution network model based on self-supervised deep learning.The model recursively reconstructed low-resolution pavement textures to a 0.1 mm·pixel-1 resolution in the driving direction.A total of 527 SMA-13 asphalt pavement textures with a resolution of 0.1 mm were prepared for training and testing.The nearest-neighbor method was used to downsample the textures by factors of 1/2,1/4,1/8,and 1/16 to simulate low-resolution textures captured at different vehicle speeds.Texture pairs with a two-fold resolution difference were used as input for the network model,and the dataset was randomly split into training and test sets in an 8∶2 ratio.The training set was augmented by small-scale segmentation.Under optimal weights,the model recursively reconstructed various low-resolution pavement textures to a 0.1 mm·pixel-1 resolution.The reconstruction quality of the super-resolution textures was evaluated using Peak Signal-to-Noise Ratio(PSNR),Structural Similarity(SSIM),and the relative error of mean texture depth,with bicubic interpolation as a comparison.By integrating bicubic interpolation with the proposed super-resolution network,this study investigated the reconstruction performance of low-resolution pavement textures under non-predefined downsampling factors.Finally,using the OOA-LightGBM algorithm,feature parameters of the high-resolution and super-resolution reconstructed textures were extracted to build a skid resistance prediction model,exploring the feasibility of predicting skid resistance based on super-resolution reconstructed textures.Results show that the proposed super-resolution network enables the vehicle-mounted laser scanning device to collect 3D pavement textures with a resolution of 0.1 mm,a PSNR greater than 30 dB,an SSIM above 0.95,and an absolute relative error in mean texture depth(MTD)of less than l％,within a speed of 128 km·h-1.For pavement textures collected at constant speeds of 16 km·h-1,32 km·h-1,and 64 km·h-1,the average coefficient of determination(R2)for skid resistance prediction was 0.808.
关键词
路面工程/路面纹理/深度学习/超分辨率网络/平均构造深度/抗滑性能Key words
pavement engineering/pavement texture/deep learning/super-resolution network/mean texture depth/skid resistance performance引用本文复制引用
出版年
2024
NETL
NSTL
万方数据