基于层次特征提取与几何模型辅助的货车故障轨边图像识别方法研究

Traditional artificial train inspection approach is inefficient, labor-intensive, and can not guarantee operating quality. The Trouble of moving Freight car Detection System has achieved the acquisition and processing of fault images, but the human eye-based fault discrimination is still less efficient. The project is planned to reveal the matching principle of the CAD model and the image pixels of train vehicles. And the parts image recognition and feature correction method based on geometric characteristics auxiliary is proposed, in order to locate the parts which it is difficult to segment only by the gray features due to the light and shelter interferences. According to the spatial distribution hierarchy of vehicle parts, the layer feature extraction method of regional spatial frequency will be explored based on Otsu multi-threshold segmentation. Then analyzing the coupled relationship among various layer features, shape features and texture features, the features selection method based on ReliefF algorithm is presented to eliminate redundant features. Based on study above, unequal weight similarity assessment and case-based reasoning adopting k-adjacent model are studied, and the hybrid reasoning mechanism for fault discrimination is designed by integrating rule-based reasoning. Following the parallel detection strategy according to passing vehicles information and fault classification, an automatic track-side image recognition system for freight car fault with high efficiency and recognition rate will be achieved ultimately. Advancing the train inspection to complete computer control, this project has a double significance to solve practical technical challenges and explore the automatic vision detection theories and methods.

传统人工列车技术检查效率低，劳动强度大，且作业质量难以保证。虽然货车故障轨边图像检测系统实现了故障图像的采集与处理，但目前以人眼为主的故障判别依旧效率较低。本项目拟揭示车辆CAD模型与图像像素的匹配原则，提出几何特征辅助的零件图像识别与特征修正方法，以定位因光照、遮挡等干扰导致仅依靠灰度特征难以分割的零件；针对车辆零部件空间分布的层次性，探索基于Otsu多阈值分割的区域空间频数层次特征提取方法，并剖析其与形状特征、纹理特征间耦合关系，研究基于ReliefF的特征选择方法以剔除冗余特征。在此基础上，研究不等权重的相似度评估与基于k-邻近模型的案例推理，并融合规则推理设计基于混合推理机制的故障判别算法；遵循基于过车信息与故障类别的并行检测策略，最终实现高效、高识别率的货车故障轨边图像自动识别系统，对推进列检向完全机控方式转变，具有解决实际技术难题和探索全自动视觉检测理论与方法的双重意义。

长期以来，列车技术检查（简称列检）基本依靠检车员“手摸、锤敲、眼看、耳听、鼻闻”完成，效率低下，劳动强度大，且作业质量易受气候、检车员素质与疲劳程度等因素影响。于是，货车故障轨边图像检测系统（Trouble of moving Freight car Detection System，TFDS）正逐步取代传统人工列检，以适应新形势下列车重载、高速、大密度运行的需要。然而，当前的TFDS大多仅仅完成货车故障图像的采集、传输以及一些预处理，而故障识别仍然依赖检车员肉眼浏览图片来完成，或由人工识别辅以计算机图像自动识别的人机结合方式，效率仍然较低。为了实现TFDS的计算机全自动图像检测与故障判别，本项目研究了几何模型辅助的零件图像识别方法、零件层次特征提取与多维故障特征的解耦理论、车辆故障知识表达与混合推理机制、基于过车分类信息的并行优化方法等。提出了基于几何模型辅助的零部件区域定位方法、基于三点迭代的聚类园拟合算法、基于改进形状上下文的零件识别算法以及组合图元识别方法，以定位因光照、噪声、遮挡等干扰而难以分割的零件，提高了故障零件识别的准确度与鲁棒性。针对车辆零部件空间分布的层次性，提出了角度与尺度混合描述子、改进高度函数的形状描述子，设计了基于这些形状描述子、几何形状特征以及不变距等零部件特征提取与匹配算法。同时，研究了基于Relief算法的多维故障特征选择方法以剔除冗余特征，针对不同类型零件选择不同的特征集与匹配算法，并结合基于规则推理与案例推理的混合推理机制以及按车型编码、部件粗定位、故障类型、识别算法逐步细化的分类处理方法，在保证零件识别正确率的基础上显著提高了检测效率。通过以上研究，设计了适应多车型、多故障并行推理与自动识别的货车故障轨边图像检测系统，以推进列检由人机结合向计算机全自动检测模式转变，同时为机器视觉理论与技术应用于其他产品质量检测和故障诊断领域提供了理论支持。