基于结构重参数化与自适应注意力的复杂路面快速识别模型  

A Fast Identification Model for Complex Pavement Based on Structural Reparameterization and Adaptive Attention

作　　者：王雪玮 李思渊 梁晓 李韶华[1,2] 郑津津 WANG Xue-wei;LI Si-yuan;LIANG Xiao;LI Shao-hua;ZHENG Jin-jin(State Key Laboratory of Mechanical Behavior and System Safety of Traffic Engineering Structures,Shijiazhuang Tiedao University,Shijiazhuang 050043,Hebei,China;Key Laboratory of Mechanical Behavior Evolution and Control of Traffic Engineering Structures in Hebei,Shijiazhuang Tiedao University,Shijiazhuang 050043,Hebei,China;School of Mechanical Engineering,Shijiazhuang Tiedao University,Shijiazhuang 050043,Hebei,China;School of Engineering Science,University of Science and Technology of China,Hefei 230026,Anhui,China)

机构地区：[1]石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,河北石家庄050043 [2]石家庄铁道大学河北省交通工程结构力学行为演变与控制重点实验室,河北石家庄050043 [3]石家庄铁道大学机械工程学院,河北石家庄050043 [4]中国科学技术大学工程科学学院,安徽合肥230026

出　　处：《中国公路学报》2024年第3期245-258,共14页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52102467,62003227,U22A20246);河北省自然科学基金项目(F2021210016,F2022210024);河北省高等学校科学技术研究项目(QN2021135);河北省省级科技计划项目(21342202D)。

摘　　要：在复杂行驶环境下快速准确地识别前方路面类型,是车辆主动控制系统及时做出预判的关键前提。针对现有方法未能兼顾精度和速度且难以在车端部署的问题,提出一种基于结构重参数化与自适应注意力的路面分类模型,可对车辆前方的沥青、水泥、冰雪、沙土、花砖、石板、湿滑等复杂路面进行快速准确的甄别。首先,构建以水平/垂直/方形/点形等多分支异构卷积为核心的特征提取骨干网络。其次,提出一种轻量高效的注意力机制,能够根据特征尺寸自适应地聚合空间上下文信息,并根据特征维度自适应地进行局部跨通道交互,使模型聚焦于高相关的路面特征。在此基础上,引入结构重参数化思想,对模型的训练周期和推理周期进行解耦,在训练时通过多分支学习获得高裕度的特征表示,而在推理时将多分支结构等价转换为直铺式单支路结构,在不牺牲模型性能的前提下获得轻量化的部署模型以及显著的推理加速。试验结果表明:提出的模型能够在复杂行驶环境下有效识别路面类型,以6.57×10^(6)的参数量取得99.14%的全场景分类精度和96.48%的新场景分类精度,同时具有496.28帧·s^(-1)的服务器推理速度和33.89帧·s^(-1)的边缘推理速度。与现有其他主流模型相比,提出的模型取得了准确性、实时性和轻量化的良好平衡,具备对复杂多变场景的高适应性,在车前路面类型识别任务上有明显优势。Fast and accurate identification of pavement types in complex driving environments is a crucial prerequisite for vehicle active control systems to generate timely predictions.In response to the problem that existing methods fail to balance accuracy and speed and are deploy-friendly,this study proposes a pavement classification model based on structural reparameterization and adaptive attention that can quickly and accurately identify complex pavements,such as asphalt,cement,snow,dirt,tile,stone,and wet asphalt.First,a feature extraction backbone network centered by multi-branch heterogenous convolutions,including horizontal/vertical/square/point kernels,was constructed.Second,a lightweight and efficient attention structure was proposed to adaptively aggregate spatial contexts depending on the feature scale and perform local cross-channel interactions depending on the feature dimension.This allowed the model to focus on highly correlated features.Accordingly,a structural reparameterization strategy was introduced to decouple the training and inference periods.During training,expressive feature representations were obtained through multi-branch learning.During inference,the multi-branch structure was equivalently transformed into a plain single-branch structure without losing performance,thereby obtaining a lightweight deploy-friendly model with a significantly high inference speed.The experimental results showed that the proposed model effectively recognized pavement types in complex driving environments.It achieves 99.14%all-scene classification accuracy and 96.48%new-scene classification accuracy with 6.57×10^(6) parameters while maintaining a high inference speed of 496.28frames per second in the server chip and 33.89frames per second in the edge chip.Compared with other models,the proposed model has higher adaptability to complex and changeable environments and achieves a better balance between efficiency,effectiveness,and lightweight.Therefore,the proposed model is significantly advantageous for pavement-typ

关 键 词：交通工程 路面类型识别 自适应注意力 结构重参数化 推理速度 轻量化模型 复杂路面 

分 类 号：U491[交通运输工程—交通运输规划与管理]