高铁接触网在途检测与跟踪算法研究 | |
陈东杰1,2 | |
2017-05-27 | |
学位类型 | 工程硕士 |
英文摘要 |
近年来我国高速铁路网建设对于产业结构升级、城市化进程加速和区域经济互联产生了助推作用。在高铁运营过程中相关部门积累了海量的视频监测数据,通过对海量时空数据的分析与处理,掌握列车运行的实时状态,可以极大地推进高铁运营管理智能化建设。在海量高铁监测数据中,弓网监测数据是主要部分。“弓”意指受电弓,“网”意指接触网,对弓网监测数据进行深度挖掘与智能分析是实现高铁智能监测的有效技术手段。目前,高铁弓网视频监测的主要方法是采用图像分析和计算机视觉技术建模,实现对零部件的检测识别、跟踪与姿态分析。然而传统的目标检测算法局限于特征描述子的设计,难以依靠人工设计出具有通用性、强鲁棒性、高精度的特征描述子。现役的高铁巡检方法主要分为人工巡检和巡检车巡检两种方式。人工巡检主要依靠工人的经验水平,虽然配备一定的辅助设备,但是辅助设备操作繁琐,费时费力且效率极低,只适用于特殊路段小范围作业;巡检车功能完善,但是由于其价格昂贵,设备数量又较少,维护成本极高,而且为了避免与列车占用轨道,巡检车适用时段狭窄,难以满足大规模检测的需要。另外,由于在高速行车途中随时可能出现不可预测的突变和干扰,现有的接触网巡检方法难以满足监测系统对于实时处理、强鲁棒性、高精度检测等日益增长的性能需求。
本论文通过对高铁接触网零部件的精细检测、跟踪以及场景文本识别,真正实现了高铁接触网的在途监测与故障预判,极大地推动了高铁运营监测智能化进程。首先,针对定位器坡度测量,采用深度卷积网络实现了实时、鲁棒、高精度的定位器检测与识别,提出了定位器骨架轮廓拟合方法,基于单目视觉实现了定位器坡度的精准测量。其次,针对弓网几何参数测量与部件缺陷检测,基于卡尔曼滤波算法提升了弓网几何参数的测量精度,并且采用深度卷积网络进行了接触网部件缺陷的检测与分析。最后,针对高铁实时定位精度辅助提升方法,采用深度卷积网络进行支柱号牌文本定位,设计了二次单字符检测算法实现了号牌字符识别,提升了高铁的实时定位精度。
本论文的主要工作和贡献如下:
1. 提出了一种实时性、强鲁棒、高精度的多尺度定位器检测与识别方法,并搭建了实际系统成功应用于定位器坡度测量模块。具体来说,首先基于深度卷积网络的弱特征微小目标检测框架,精准定位出目标最小包络矩形区域,然后基于Hough直线变换求得定位器轮廓的最优拟合直线段,最后基于单目视觉结合模型的先验知识,求出定位器的三维空间坐标值,实现了定位器坡度的非接触式精准测量。
2. 提出了弓网几何参数测量与部件故障检测方法。第一部分内容是针对接触线等特殊目标的检测与跟踪,提出了基于卡尔曼滤波的“拉索导线”协同跟踪机制,算法克服了隧道内等复杂背景下接触线的像素显著性不强这一先天性缺陷,在加载补光灯的成像环境下,提升了接触线的识别与跟踪精度。第二部分内容是针对接触网零部件的故障检测,提出了基于深度卷积网络的零部件缺陷检测算法,实现了接触网零部件故障的精细检测。
3. 提出了一种高铁实时定位辅助方法。首先基于深度卷积网络实现铁路沿线支柱号牌的精准定位,然后提取号牌区域的图像部分,最后结合场景文本识别算法实现了支柱号牌编号的识别。由于接触网支柱号牌编号符合一定的编排规则,所以通过对铁路沿线的支柱号牌编号识别即可推断出列车所处的实时位置。同时我们利用特定场景下的先验知识进一步优化了接触网号牌识别精度。 ;
In recent years, China High-speed Railway Network plays an important role in the industry structure upgrade, urbanization acceleration and regional economic interconnection. Meanwhile, with continuous operation of high-speed railway, we have accumulated massive video monitoring data, including the time and space data. It is essential for us to master the real-time state of the running train through the “Bow-Net” monitoring data analysis, which can greatly promote the high-speed railway intelligent management system. Among the whole data, Bow-Net system monitoring data is the main part. “Bow” refers to the pantograph, “Net” refers to the catenary. At present, general method of high-speed railway network video surveillance is conducted by model construction, integrating all the prior knowledge, digital image processing and computer vision technology into practical process, in order to achieve target detection and recognition, target tracking, and even posture analysis. However, developed target detection algorithms are limited by the design of artificial feature descriptors, it is difficult to rely on artificial design a feature descriptor with the characteristics of generality, robustness, and high-accuracy. Developed monitoring methods can be split into two groups—manual inspection and vehicle inspection. Manual inspection relies mainly on the workers experience level, although it has certain auxiliary equipment, but it has disadvantages such as time consuming and low efficiency, so it is suitable for special sections for small scale operation. Functional vehicle inspection device can hardly satisfy the requirement of real-time processing because of its expensive, less equipment, high maintenance costs, and its applicable period is very narrow in order to avoid conflicting with the running train. But even worse, due to captured images may appear unpredictable mutation and interference, traditional target detection methods can hardly satisfy the requirement of real-time processing,robustness, and high-accuracy detection.
In this paper we achieve online multi-target detection, tracking, and even scene text recognition of high-speed railway catenary parts, which plays an important role on High-speed Railway catenary intelligent monitoring system. First of all, we achieve the real-time, high precision detection and recognition of catenary locator based on deep convolution network, furthermore, we complete the locator slope precise measurement through skeleton outline fitting method under monocular vision framework. Second, we adopt the optimize model based on Kalman filter to improve cable and wire line recognition and tracking accuracy, in order to separate them when complicated backgrounds emerge, such as in tunnel, under bridge, and in station yard. Finally, in order to improve high-speed railway positioning accuracy, we design single character twice detection algorithm to achieve pole plate text recognition, and our method greatly improves the high-speed railway localization accuracy.
In this paper, our main work and contributions are as follows:
1. We present a robust and real-time, high precision of target detection and identification method and the application system, this solution method is successfully applied in the “Bow-Net” component detection and fault analysis system. We achieve tiny and feature-weak target detection based on deep convolution network. After positioning an outline of target precisely, we optimize the optimum fitting line segments based on Hough line transform, and we complete the locator slope non-contact precise measurement based monocular vision framework finally.
2. We propose novel detection methods of “Bow-Net” geometrical parameters and its component breakdown. The first part is that we put forward a collaborative tracking method based on Kalman filter, which realizes identification and tracking of the cable and wire lines of high-speed railway in complicated background and extreme environment. This method overcomes the birth defects of cable and wire lines captured in tunnel, which pixels saliency is not obvious enough. The second part is the fault detection of the catenary component. We use the target detection algorithm based on deep convolution network to realize the precise detection of catenary component breakdown.
3. We propose a real-time High-speed Railway positioning auxiliary methods, using computer vision technology, through deep convolution network for high-precision localization of pole plate, and then we achieve real-time pole plate number recognition by Scene Text Recognition algorithm and realize real-time positioning due to the pole plates conform to certain rules. Meanwhile, we adopt priori knowledge to refine the text recognition precision. |
关键词 | 高铁接触网 目标检测 目标跟踪 深度学习 场景文本识别 |
学科领域 | 模式识别与智能系统 |
文献类型 | 学位论文 |
条目标识符 | http://ir.ia.ac.cn/handle/173211/14765 |
专题 | 毕业生_硕士学位论文 |
作者单位 | 1.中国科学院自动化研究所 2.中国科学院大学 |
第一作者单位 | 中国科学院自动化研究所 |
推荐引用方式 GB/T 7714 | 陈东杰. 高铁接触网在途检测与跟踪算法研究[D]. 北京. 中国科学院大学,2017. |
条目包含的文件 | ||||||
文件名称/大小 | 文献类型 | 版本类型 | 开放类型 | 使用许可 | ||
硕士毕业论文_陈东杰final.pdf(6140KB) | 学位论文 | 限制开放 | CC BY-NC-SA |
个性服务 |
推荐该条目 |
保存到收藏夹 |
查看访问统计 |
导出为Endnote文件 |
谷歌学术 |
谷歌学术中相似的文章 |
[陈东杰]的文章 |
百度学术 |
百度学术中相似的文章 |
[陈东杰]的文章 |
必应学术 |
必应学术中相似的文章 |
[陈东杰]的文章 |
相关权益政策 |
暂无数据 |
收藏/分享 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论