基于单目视频流的前方车辆检测与识别

发布时间：2018-05-04 01:41

本文选题：图像处理 + 车道线提取　；参考：《吉林大学》2017年硕士论文

【摘要】：智能驾驶车辆在社会进步和科技发展中成为主要话题,为传统的汽车解决许多棘手的安全问题。因此,智能车辆障碍检测技术的研究越来越受到重视,而可靠地检测前方车辆对于智能驾驶具有重大的研究意义。本文是通过在电动车前方架设CCD摄像头来获取道路信息,信息处理过程包括图像处理、车道线检测、感兴趣区域划分,对称性及几何特征识别前方车辆。基于CCD摄像头的前方车辆检测便可以时刻监测道路前方环境,对构成潜在安全隐患的前方进行检测。论文主要研究内容包括四个方面:1.单目视频流的采集及图像预处理。介绍了经CCD摄像头采集到的视频流是由连续的静态帧组成的,并在检测前方车辆前需要对图像预处理。图像预处理过程主要包括图像的灰度化及二值化、图像滤波、图像的边缘检测,其中选择基于Canny算子检测图像边缘,实验表明二阶微分算子在边缘检测中可以有效提取图像中边缘信息。2.感兴趣区域的提取。对前方车辆进行检测与识别时,首先需要确定前方车辆的候选区域,本文针对改进的Hough变换对车道线进行识别和标注,利用区域分割方法标注感兴趣区域,即感兴趣区域含有待检测的前方车辆,并通过划分结果验证系统车道线检测的结果。3.车辆检测识别。在前方车辆检测的算法中,本文论述了基于对称性及提取车辆特征点等算法,重点在于特征点提取中的改进的Harris角点检测,在实时性取得突破,在除去边界点检测方面加以创新并作为改进算法的核心思想,为下文对称性及几何特征匹配车辆阶段奠定基础,确保了车辆检测结果的准确性。4.基于时空上下文的视频流车辆跟踪。确定静态帧的车辆检测结果后,需要对视频流中的车辆进行跟踪,以此作为智能驾驶系统中保持安全距离必不可少的部分。本文用时空上下文算法在视频流序列中的车辆追踪,满足了围绕视频序列前后帧的空间上下文特性。前方车辆检测技术可以有效减少交通事故并保障车辆安全驾驶,障碍检测不仅应用于交通领域,同时在工业应用、科学探测、救灾抢险、国防军事等领域也有着广泛的应用前景。
[Abstract]:Intelligent driving vehicle has become the main topic in the progress of society and the development of science and technology, solving many thorny safety problems for traditional cars. Therefore, more and more attention has been paid to the research of intelligent vehicle obstacle detection technology. In this paper, CCD camera is set up in front of electric vehicle to obtain road information. The process of information processing includes image processing, lane detection, region of interest division, symmetry and geometric features recognition. The front vehicle detection based on CCD camera can monitor the road front environment at all times and detect the potential safety hidden danger. The main content of this paper includes four aspects: 1. Monocular video stream acquisition and image preprocessing. This paper introduces that the video stream collected by CCD camera is composed of continuous static frames, and the image preprocessing is needed before detecting the vehicle in front. The process of image preprocessing mainly includes grayscale and binarization of image, image filtering, edge detection of image. Among them, Canny operator is chosen to detect image edge. Experiments show that the second order differential operator can effectively extract the edge information from the image in edge detection. Extraction of regions of interest. In order to detect and identify the vehicle in front, we need to determine the candidate area of the vehicle in front. In this paper, the lane line is identified and marked by the improved Hough transform, and the region segmentation method is used to mark the region of interest. In other words, the region of interest contains the vehicle in front of the vehicle to be detected, and the result of lane line detection is verified by the partition result. 3. Vehicle detection and identification. In the forward vehicle detection algorithm, this paper discusses the algorithms based on symmetry and vehicle feature points, the emphasis is on the improved Harris corner detection in feature point extraction, which makes a breakthrough in real-time. As the core idea of the improved algorithm, it can provide the foundation for the following phase of symmetry and geometric feature matching vehicle, and ensure the accuracy of the vehicle detection results. 4. Video stream vehicle tracking based on temporal and spatial context. After determining the vehicle detection results of the static frame, it is necessary to track the vehicle in the video stream as an essential part of the intelligent driving system to keep a safe distance. In this paper, the spatio-temporal context algorithm is used to track the vehicle in the video stream sequence, which satisfies the spatial context characteristics around the frame before and after the video sequence. The forward vehicle detection technology can effectively reduce traffic accidents and ensure the safe driving of vehicles. Obstacle detection is not only applied in the field of transportation, but also in industrial applications, scientific detection, disaster relief and rescue. National defense military and other fields also have a wide range of applications.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】