基于任意四边形拟合的轨道分割算法  

Track Segmentation Algorithm Based on Arbitrary Quadrilateral Fitting

作　　者：郭浩 齐咏生[1,2,3] 张嘉英[1,2,3] 马然[1,2,3] 刘利强[1,2,3] GUO Hao;QI Yongsheng;ZHANG Jiaying;MA Ran;LIU Liqiang(School of Electric Power,Inner Mongolia University of Technology,Hohhot 010080,China;Engineering Research Center of Large Energy Storage Technology of Ministry of Education,Inner Mongolia University of Tcehnology,Hohhot 010080,China;Center for Intelligent Energy Technology andEquipment Engineering,Inner Mongolia University of Technology,Hohhot 010080,China)

机构地区：[1]内蒙古工业大学电力学院,内蒙古呼和浩特010080 [2]内蒙古工业大学大规模储能技术教育部工程研究中心,内蒙古呼和浩特010080 [3]内蒙古工业大学内蒙古自治区高等学校智慧能源技术与装备工程研究中心,内蒙古呼和浩特010080

出　　处：《铁道学报》2025年第1期112-122,共11页Journal of the China Railway Society

基　　金：国家自然科学基金(62363029);内蒙古自治区科技计划(2020GG0283,2021GG0256);内蒙古自治区自然科学基金(2020MS05029,2021MS06018)。

摘　　要：精准高效地完成轨道分割与障碍物检测对于智能行车具有重要意义。现有深度学习算法对轨道区域分割与障碍物识别区分不明确,复杂区域极易出现多轨融合、识别率低等情况。为此,提出一种基于任意四边形拟合的轨道分割框架,实现轨道分割任务;将其嵌入YOLOv5网络中,设计出一种全新的轨道交通双视觉任务网络(YOLOv5-DVT),该网络可根据轨道的不同特征,使用任意四边形切分轨道;经过多边形角点排序、自适应正样本匹配、相对重心位置编解码及顺序预测约束等环节完成对任意四边形轨道的预测,利用多边形寻迹拟合算法实现对轨道区域的恢复与分割;通过设计双任务结构,采用端到端的轨道分割与障碍物检测并行训练策略实现同步分割与障碍物识别,提升推理速度。采用自建数据集对该算法进行验证,试验结果表明:与经典分割算法相比,本文方法在复杂轨道的分割中更加清晰准确,其双任务精度分别达到95.10%、93.51%,推理速度达32 FPS,具备实际场景的应用价值。In the field of rail transportation safety,accurate and efficient completion of track segmentation and obstacle detection is of great significance for intelligent train operation.However,the unclear differentiation of existing deep learning algorithms between track area segmentation and obstacle task recognition may cause multi⁃track fusion and low recognition rate in complex areas.Therefore,a track segmentation framework was proposed based on arbitrary quadrilat⁃eral fitting,which was then embedded into YOLOv5 network to design a new network of dual vision task for rail transpor⁃tation(YOLOv5⁃DVT).The network can use arbitrary quadrilateral to slice the track according to different features of the track.The prediction of atypical quadrilateral was completed through polygon corner point sorting,adaptive positive sample matching,relative center of gravity position coding and decoding and sequential prediction constraints.The re⁃covery and segmentation of the track region were realized by using polygon tracking and fitting algorithm.Finally,the in⁃ference speed was improved by designing a dual⁃task structure and adopting an end⁃to⁃end track segmentation and obsta⁃cle detection parallel training strategy to achieve simultaneous segmentation and obstacle recognition.The algorithm was verified by using a self⁃built dataset.The experimental results show that compared with the classical segmentation algo⁃rithm,this method is clearer and more accurate in the segmentation of complex tracks,with dual⁃task accuracy reaching 95.10%and 93.51%respectively,and the inference speed reaching 32 FPS,which has practical application value in practical scenarios.

关 键 词：任意四边形检测 轨道分割 障碍物检测 深度神经网络 

分 类 号：U213.2[交通运输工程—道路与铁道工程]