基于前向运动视频的计算机视觉检测技术研究及应用

发布时间：2018-07-26 18:57

【摘要】：前向运动的视频拍摄作为一种移动式的场景获取方式,由于其视野宽阔,空间覆盖面广,已经被广泛应用于移动式的场景监控和目标检测任务中。然而,随着视频资源的数据量不断增加,许多算法由于计算复杂性逐渐无法满足检测任务的实时性要求,而且庞大的视频数据量也为数据的存储和检索造成了困难。本文的研究基于摄像机前向运动拍摄的视频数据,从理论和应用两个方面展开论文的工作。首先提出了基于检测区域几何结构的全景环带采样方法,并构建快速的前向运动视频的全景拼接算法,将海量的视频数据进行无损的信息抽取,得到了轻量级的全景图格式,不仅降低了视频数据的存储和访问开销,而且将视频转化为一种更适合于人工检视或计算机分析处理的形式。之后,以铁路轨道状态和护栏缺损检测为应用研究背景,提出了相应的基于全景图的自动化视觉检测算法。本论文的创新工作如下：1.提出了一种基于检测区域几何结构的全景采样模型。不同于以往的视频拼接模型和图像对齐方法,本文提出的基于检测区域几何结构的全景采样方法可以快速简洁地从前向运动视频中生成全景图。在相机的内部参数已标定的情况下,仅仅依靠摄像机的运动信息和空间场景的几何结构先验,完成了拼接区域的构造和对齐,没有执行耗时的图像匹配和复杂的光流计算,实现了铁路环境视频的实时拼接,从而确保了基于全景图的后续视觉检测的实时性。2.提出了一种基于“双缝投影”的单目立体全景成像方法。首先构建了立体全景采样的“双缝投影”模型从每一帧图像中抽取一对具有不同视角的拼接条带,用来生成两幅具有显著视差的全景图；然后分析了立体全景图成像的原理来推导全景图深度的计算公式；最后利用全景图拼接方法生成了立体全景图像对,并基于局部立体匹配算法来估算“像位差”,从而获得空间场景的深度信息。立体全景图具有更广阔的视野和更强的真实感,场景的深度信息也为人工目视检测或计算机自动识别提供了更加丰富的决策信息。3.提出了一种基于轨道全景图(Rail track panorama)的钢轨健康状态的自动感知方法。在物理空间中,由于轨道不良状态而引起的列车晃动,势必引起图像空间中轨道全景图(RTP)的钢轨形态的变化。首先通过分析列车晃动下的几何成像模型,推导出“轨道不良状态—列车晃动—钢轨图像失真”之间的联系。之后利用图像阈值分割和形态学滤波的方法从轨道全景图上提取钢轨轮廓图像,并分析钢轨轮廓图像的失真变化来反演出轨道的健康状态。4.提出了一种基于护栏全景图(Fence panorama)的护栏缺损的自动化检测方法。一旦获得整个连续护栏的全景图,那么就可以首先利用阈值分割的方法自动提取护栏中竖直栏杆的位置,使其与背景图像分离：其次根据分析相邻栏杆之间的距离进行护栏缺损判别。本文提出了一种“基于MVG三维直方图的最大熵阈值分割方法”实现了栏杆位置的定位,并基于行程编码的思想将护栏全景图进行编码,压缩的编码格式包含了护栏的全部位置信息,大幅降低了存储开销,是一种有效的表示方法和存储格式。同时也设计了相应的解码算法从编码中恢复护栏的位置并实现了护栏的缺失检测。
[Abstract]:As a moving scene acquisition mode, forward motion video capture has been widely used in mobile scene monitoring and target detection tasks due to its wide field of vision and wide space coverage. However, with the increasing amount of data in video resources, many algorithms are gradually unable to meet the detection tasks due to the complexity of computing. The real time requirement and the huge amount of video data have also caused difficulties in the storage and retrieval of data. This study is based on two aspects of the theory and application of video data taken from the forward motion of the camera. First, a panoramic band sampling method based on the detection of the geometric structure of the region is proposed, and the rapid development of the method is proposed. The panoramic splicing algorithm for forward motion video is used to extract massive video data from nondestructive information and obtain a lightweight panorama format. It not only reduces the storage and access overhead of video data, but also transforms video into a form that is more suitable for artificial viewing or computer analysis. As the application research background, the automatic visual detection algorithm based on panorama is proposed. The innovation work of this paper is as follows: 1. a panoramic sampling model based on the detection region geometry is proposed. The proposed method is based on the detection of the video mosaic model and the image alignment method. The panorama of the regional geometric structure can quickly and succinctly generate panoramic images from the previous motion video. In the case of the calibration of the camera's internal parameters, the construction and alignment of the stitching region is completed only by the motion information of the camera and the geometric structure of the space scene, and the time-consuming image matching and complex is not performed. The real-time mosaic of the railway environment video is realized by the miscellaneous optical flow calculation, which ensures the real-time.2. based on the follow-up vision detection based on the panorama. A single stereo panoramic imaging method based on the "double slit projection" is proposed. The splice strips with different perspectives are used to generate two panoramic images with significant parallax, and then analyze the principle of stereotactic imaging to derive the calculation formula for the depth of the panorama. Finally, the panoramic image pair is generated by the panoramic image stitching method, and the "image bit difference" is estimated based on the local erect matching algorithm. The depth information of the space scene is obtained. The stereoscopic panorama has a broader vision and a stronger sense of authenticity. The depth information of the scene provides a more abundant decision information for artificial visual inspection or computer automatic recognition..3. proposes an automatic perception side for the health of rail based on the Rail track panorama. Method. In the physical space, the train sloshing due to the bad state of the track will cause the change of the rail form of the track panorama (RTP) in the image space. First, by analyzing the geometric imaging model under the sloshing of the train, the connection between the bad state of the train and the distortion of the rail image in the train is deduced. Such as threshold segmentation and morphological filtering, the rail profile is extracted from the track panorama, and the distortion of the rail profile is analyzed to reverse the health state of the track. An automatic test method based on the guardrail defect based on the fence panorama (Fence panorama) is proposed. Once the panorama of the whole continuous guardrail is obtained. Then, the position of the vertical railing in the guardrail can be automatically extracted by the method of threshold segmentation, so that it can be separated from the background image. Secondly, according to the analysis of the distance between adjacent rails, a kind of "maximum entropy threshold segmentation method based on MVG 3D histogram" is proposed to implement the railing position. It encodes the panoramic view of the guardrail based on the idea of stroke coding. The compression coding format contains all the location information of the guardrail, greatly reduces the storage cost. It is an effective representation method and storage format. At the same time, the corresponding decoding algorithm is designed to restore the position of the guardrail from the code and realize the lack of the guardrail. Misdetection.
【学位授予单位】：北京交通大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41;R339.14

【相似文献】