列车前方轨道识别算法的设计与实现
发布时间：2019-01-09 11:43
【摘要】：近年来,随着城市轨道交通的快速发展,地铁、轻轨等列车的行车安全变得日益重要。障碍物检测系统通过辅助驾驶员排查列车前方轨道情况,提高了列车的行车安全性。对于障碍物检测系统,高效准确地识别出列车前方轨道至关重要。本文基于列车车载前置摄像头采集到的图像,完成了对列车前方轨道的识别。识别过程主要分两步,首先是近距离轨道识别,然后是根据近距离轨道识别的结果得到种子点,以改进的引入方向的种子区域增长方法完成远距离轨道识别。对于近距离里轨道识别,本文采用的是现有的基于曲率映射图的轨道识别算法,并作了进一步的改进。其中,曲率映射图类似于模板匹配中的模板,不过更为细致。本文的改进主要体现在曲率映射图的创建上,在创建曲率映射图之前,建立了列车与不同曲率的理想轨道之间的位置关系。然后根据相机的内参及其相对列车的位置与姿态得到图像上每个像素点对应的理想轨道的曲率,即曲率映射图。整个过程不仅计算方便,而且得到的曲率映射图精度高。另外,本文在根据曲率映射图和输入图像的梯度图选取最匹配的理想曲率时,也作了一定的改进。对于远距离轨道识别,本文提出了基于局部梯度信息的轨道识别算法,该方法一共分四步。首先根据图像梯度,设计一个度量,衡量某一区域内的图像与实际轨道图像的相似度。然后根据近距离轨道识别的结果,得到初始的种子点(包括位置和方向)。接着在当前初始种子点的邻域内搜索一个最佳的位置和方向(相似度最高,并满足一定约束的)作为当前最佳种子点,并由当前最佳种子点延伸到下一个初始种子点。最后重复种子延伸过程,并联合左右钢轨一同进行,直到找不到满足一定约束的最佳种子点,从而完成远距离轨道的识别。此外,本文还提出了基于两条已知间距的平行线的相机外参标定算法。该方法根据曲率为0,坡度不变的直线型轨道在图像上的位置标定得到相机外参,标定过程方便快捷。
[Abstract]:In recent years, with the rapid development of urban rail transit, the safety of trains such as subway and light rail has become increasingly important. The obstacle detection system improves the safety of the train by assisting the driver to check the track in front of the train. For the obstacle detection system, it is very important to identify the track in front of the train efficiently and accurately. Based on the image collected by the front camera of the train, the recognition of the track in front of the train is completed. The recognition process is mainly divided into two steps: first, the close orbit recognition, and then the seed points are obtained according to the results of the close orbit recognition, and the improved seed region growth method with introduced direction is used to complete the long distance orbit recognition. For the near distance orbit recognition, this paper uses the existing curvature mapping graph based orbit recognition algorithm, and makes further improvements. Among them, curvature map is similar to template in template matching, but more detailed. The improvement of this paper is mainly reflected in the creation of the curvature mapping graph. Before the curvature mapping graph is created, the position relationship between the train and the ideal track with different curvature is established. Then the curvature of the ideal orbit corresponding to each pixel in the image is obtained according to the camera's internal parameters and the relative position and attitude of the train, that is, the curvature map. The whole process is not only easy to calculate, but also has high precision of curvature map. In addition, this paper also makes some improvements in selecting the most suitable ideal curvature according to the curvature mapping graph and the gradient map of the input image. For long distance orbit recognition, an algorithm based on local gradient information is proposed, which is divided into four steps. Firstly, according to the image gradient, a measure is designed to measure the similarity between the image in a certain region and the actual track image. Then the initial seed points (including position and direction) are obtained according to the results of close orbit recognition. Then we search the neighborhood of the current initial seed point for an optimal position and direction (with the highest similarity and satisfying certain constraints) as the current best seed point and extend from the current best seed point to the next initial seed point. Finally, the process of seed extension is repeated and combined with the left and right rails, until the best seed point can not be found to meet certain constraints, thus the identification of the long distance orbit is completed. In addition, a camera external parameter calibration algorithm based on two known parallel lines is proposed. In this method, the camera parameters can be obtained according to the position of the linear track with a curvature of 0 and a constant slope on the image, and the calibration process is convenient and fast.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U284.48;TP391.41
本文编号：2405576
[Abstract]:In recent years, with the rapid development of urban rail transit, the safety of trains such as subway and light rail has become increasingly important. The obstacle detection system improves the safety of the train by assisting the driver to check the track in front of the train. For the obstacle detection system, it is very important to identify the track in front of the train efficiently and accurately. Based on the image collected by the front camera of the train, the recognition of the track in front of the train is completed. The recognition process is mainly divided into two steps: first, the close orbit recognition, and then the seed points are obtained according to the results of the close orbit recognition, and the improved seed region growth method with introduced direction is used to complete the long distance orbit recognition. For the near distance orbit recognition, this paper uses the existing curvature mapping graph based orbit recognition algorithm, and makes further improvements. Among them, curvature map is similar to template in template matching, but more detailed. The improvement of this paper is mainly reflected in the creation of the curvature mapping graph. Before the curvature mapping graph is created, the position relationship between the train and the ideal track with different curvature is established. Then the curvature of the ideal orbit corresponding to each pixel in the image is obtained according to the camera's internal parameters and the relative position and attitude of the train, that is, the curvature map. The whole process is not only easy to calculate, but also has high precision of curvature map. In addition, this paper also makes some improvements in selecting the most suitable ideal curvature according to the curvature mapping graph and the gradient map of the input image. For long distance orbit recognition, an algorithm based on local gradient information is proposed, which is divided into four steps. Firstly, according to the image gradient, a measure is designed to measure the similarity between the image in a certain region and the actual track image. Then the initial seed points (including position and direction) are obtained according to the results of close orbit recognition. Then we search the neighborhood of the current initial seed point for an optimal position and direction (with the highest similarity and satisfying certain constraints) as the current best seed point and extend from the current best seed point to the next initial seed point. Finally, the process of seed extension is repeated and combined with the left and right rails, until the best seed point can not be found to meet certain constraints, thus the identification of the long distance orbit is completed. In addition, a camera external parameter calibration algorithm based on two known parallel lines is proposed. In this method, the camera parameters can be obtained according to the position of the linear track with a curvature of 0 and a constant slope on the image, and the calibration process is convenient and fast.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U284.48;TP391.41
【相似文献】
相关博士学位论文 前1条
1 陈新兴;随机映射图的局部性质[D];复旦大学;2007年
相关硕士学位论文 前3条
1 王云泽;列车前方轨道识别算法的设计与实现[D];浙江大学;2017年
2 张颂;基于光子映射图真实感图像生成系统及参与介质绘制研究[D];天津大学;2004年
3 詹玲;基于平板PET晶体区域映射图的识别算法研究[D];华中科技大学;2007年
，本文编号：2405576
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2405576.html
