【作者】 李龙；

【导师】 王耀南；

【作者基本信息】 湖南大学 ， 控制科学与工程， 2016， 硕士

【摘要】 基于视觉的实时车辆检测与跟踪是智能交通领域的一个重要研究课题,在车辆辅助驾驶、危险报警等方面有着很好的应用前景。前方车辆的检测与跟踪用于准确的检测前方是否有车辆,并为车辆辅助驾驶系统提供有效地驾驶环境信息。本文主要研究单目视觉的前方车辆检测和跟踪算法,通过对比分析国内外的车辆检测和跟踪算法,提出了一种基于改进ACF(Aggregated Channel Features,基于聚集通道特征的目标检测器)算法的前方车辆检测算法和一种基于多子模板的NCC(Normalized Correlation Coefficient,基于归一化互相关系数的图像相似性度量)匹配车辆跟踪算法,并实现了前方车辆的检测与跟踪软件系统。主要研究内容有两点:(1)前方车辆检测算法研究。本文提出了一种基于改进ACF算法的车辆检测算法,该方法能够实时检测复杂的城市道路上的前方车辆。通过分析前方车辆的特点,将车尾的垂直边缘特征引入到ACF特征中,与原有的10通道特征共同组成新的11通道特征,并用改进的ACF特征检测前方车辆,分类器选用Adaboost,得到初始检测结果,然后对初始检测结果进行NMS,去除多余检测窗口,再采用车尾灰度对称性验证检测窗口,最后结合ACF检测结果与车尾灰度对称性验证结果得出最终检测结果。算法实时性好,检测精度高,适用于复杂的城市道路上前方车辆检测。(2)前方车辆跟踪算法研究。本文提出一种基于多子模板的NCC匹配跟踪算法,算法通过在目标模板中选取若干个矩形小块作为子模板,再剔除标准差小于一定阈值的子模板,然后建立搜索图像金字塔,用NCC算法计算每个子模板与金字塔图像的归一化互相关系数矩阵,最后,综合所有子模板与金字塔图像的归一化互相关系数矩阵建立最终的跟踪结果。算法不仅跟踪精度高、有较好的实时性能,还具有抗部分遮挡和光照变化的优点。在软件系统实现中,本文采用Visual Studio 2013+Open CV 2.4.9作为编程开发环境,开发出了前方车辆检测与跟踪系软件,软件主要包括图像采集模块、车辆检测模块、车辆跟踪模块以及决策模块。对实验结果分析表明:本文的车辆检测和跟踪算法能够准确的检测并跟踪前方车辆,并且能够满足实时检测与跟踪的要求。更多还原

【Abstract】 Real-time vehicle detection and tracking is an important researching field in intelligent traffic. It is important for auxiliary driving system and danger warning for drivers. The forward vehicle detection and tracking system provides cues about the number of the cars ahead. It also provides effective environment information for auxiliary driving systems.In this paper, we are targeted for an effective forward car detection and tracking algorithm based on monocular vision. An improved ACF(Aggregated Channel Features, Aggregated Channel Features based object detector) algorithm is proposed for forward car detection. A multiple templates based car tracking algorithm based on NCC(Normalized Correlation Coefficient, A traditional image similarity measurement based on Normalized Correlation Coefficient) matching is also proposed to tackle the tracking problem. Experimental results demonstrated the effectiveness of the proposed algorithms on real-world data. Finally, we implemented a forward car detection and tracking system by combining the proposed detection and tracking algorithms.The major part of our research are two folds:(1) An effective forward-car detection algorithm. We propose a car detection algorithm based on the ACF algorithm. The proposed algorithm can detect cars in complex urban road networks in real-time. By analyzing the forward car images, we found the vertical edges on the back of cars are significant indications of car existences. We thus incorporate that feature into ACF, together with the original 10 feature channels, getting an 11 channel feature image representation. We adopts the Adaboost algorithm for image classification and generate initial detections by sliding a window in a feature pyramid, where each feature level is computed using the 11 feature channels. The initial detections are filtered by non-maximum-suppression(NMS). Then a new symmetry measure is calculated to verify true cars. The proposed method achieves both accuracy and efficiency in complex urban environments.(2) An effective vehicle tracking algorithm. We propose a multiple templates based NCC matching based tracking algorithm. The algorithm selects several sub-regions in the target area and builds corresponding templates. For effectiveness, the standard deviations of the templates are required to be greater than a certain threshold. For tracking, the searching image is built a pyramid and convolved with all the templates at different scales. The convolution operation here is rather a NCC matching process which computes NCC response by sliding a template in an image. Finally, all the template responses are integrated to give a final decision about target position and scale. The proposed algorithm is not only accurate, it also can effectively handle the tracking challenges brought by partial occlusion and illumination variation. Besides, the tracking algorithm runs in real time.We implemented the proposed detection and tracking algorithms in C++ by Visual Studio 2013 and Open CV 2.4.9. The resulting forward car detection and tracking software system consists of several modules including image acquisition, forward car detection, vehicle tracking and decision process. Experiments with real world images show that the proposed algorithms can detect and track forward cars effectively, and suggest for real-time detection and tracking applications.更多还原