基于机器视觉检测的铁路轨道表面缺陷快速识别与分类方法研究

The running rails in service will cause various defects, such as rail damage, broken sleepers, fasteners loss and so on, due to the traffic load as well as the natural factors, which can be a threat to the safety of the train running. Because of the long rail line and complex terrain, manual detecting is dangerous and difficult. The technology of machine vision detection based on image analysis is one of the hot topics in the field of the non-contact detection of the current track defects currently. However, in the practical applications, the factors of illumination, sheltering and complex and various environments exist. Thus, the quality of track images obtained from visual means is ragged, which will have the high requirements for the following understanding of surface information of track defects based on image analysis. Therefore, the efficient sparse representation model based on multi-scale geometric analysis and vision images of track defects will be established according to the characteristics of images and videos of track defect and on the base of the surface detection technologies of track defect at home and abroad.The research on the efficient and fast image registration, the division, classification and recognition of track defect will be done under the above model. Meanwhile, the classifier based on compressed sensing will be built to broaden the strict requirements for feature extraction and resolve the problems of correct detection of rail fastening defects with the sheltering factor.

运营铁路轨道在使用过程中,由于行车载荷以及自然因素的作用,会使轨道等产生各种缺陷（如钢轨损伤、轨枕破损、扣件丢失等），对列车运行的安全构成威胁，因轨道路线长,地形复杂,人工检测比较危险和困难。基于图像分析的机器视觉检测技术是当前轨道缺陷非接触检测的热门研究方法之一。实际应用中存在光照、遮挡及环境复杂多变等因素的影响，基于视觉手段获取的轨道图像质量参差不齐，对后续进行的基于图像分析的轨道缺陷表面信息理解提出了很高的要求。为此在研究国内外轨道缺陷表面检测技术的基础上，根据轨道缺陷图像及视频的特点，建立轨道缺陷视觉图像的多尺度几何分析的高效稀疏表示模型，在此模型上研究高效快速的图像配准、轨道缺陷分割和缺陷识别及分类算法，建立基于压缩感知理论的分类器，以放宽对特征提取的严苛要求，解决存在遮挡干扰下的轨道扣件缺陷正确检测问题。

铁路服务于国家的经济建设和人民的物质文化生活，随着我国铁路向高速和重载方向发展，对钢轨可靠性检测提出了严格要求。传统代替人工巡检的方法存在检测效率低、准确率不高等不足。近年来，计算机视觉技术以其快速、灵敏、准确性高等优点逐渐应用到钢轨检测领域。申请人通过对大量钢轨图像的特征分析，提出了基于灰度突变方法对钢轨轨面和扣件区域进行了提取，考虑自然光照和图像采集过程中噪声的影响，改进了Retinex方法对钢轨表面图像进行了增强，基于双边滤波器设计了背景平滑滤波器，利用局部灰度和梯度变化信息自适应调整不同特征区域平滑程度。并运用背景差分方法通过差分图像阈值分割，利用连通区域标记方法，实现了钢轨表面缺陷检测。利用开关型中值滤波和基于图像梯度幅值的改进Canny边缘检测方法，对扣件边缘特征进行自适应图像增强，结合扣件弹条稳定的内外边缘轮廓特征，利用基于曲线特征投影的模板匹配实现了扣件缺失的实时检测。经过现场实验证明，项目组提出的轨面缺陷分割方法和扣件缺失检测方法均能满足钢轨缺陷检测要求。