基于单目视觉的智能行车预警系统的技术研究

发布时间：2018-03-25 15:01

本文选题：车辆识别　切入点：视觉检测　出处：《南京理工大学》2016年硕士论文

【摘要】：本文研究了基于单目视觉的行车预警系统,分别在白天车辆检测、夜间车辆检测以及车辆跟踪三大技术领域内对国内外各种研究成果进行了分析对比,根据不同算法的特点以及在实际场景中遇到的问题设计了相应的改进算法。在白天场景下的车辆检测中,针对传统的基于Hough变换的车道线检测算法不能提取弯道车道线的缺点,设计了基于对称搜索的四次多项式车道线拟合算法,减小了拟合误差。针对直接将疑似车辆目标结果输入到SVM分类器进行分类的方法效率低的特点设计了基于PCA降维算法的验证算法,即通过提取初步检测结果的特征向量简化分类器的输入提高算法的检测效率。在夜间车辆检测中,针对传统算法通过单一颜色阈值来分割车辆尾灯会误检测出其他发光背景的情况,分别建立了尾灯红色光晕以及白色中心的HSV颜色模型,通过分别提取这两种颜色来识别出车辆尾灯,并在获得车辆尾灯初步检测的基础上利用证据融合算法来验证检测到的车辆尾灯。在车辆跟踪过程中,设计了基于运动模型和跟踪队列的车辆跟踪算法,通过运动模型预估车辆目标在下一帧中的位置,利用图像匹配算法在预估位置周围搜索车辆目标,用搜索结果对运动模型进行修正,并通过建立目标的跟踪队列来完成对车辆的跟踪。通过对实际道路场景上拍摄的视频进行处理的结果表明,在Windows 10操作系统以及3.2GHz CPU的环境下,采用本文设计的车辆检测与跟踪方法能够准确的检测出前方车辆并且能够达到每秒20帧的检测速度,能够满足实时性的要求。
[Abstract]:In this paper, the vehicle warning system based on monocular vision is studied. The research results are analyzed and compared in three technical fields: daytime vehicle detection, night vehicle detection and vehicle tracking. According to the characteristics of different algorithms and the problems encountered in the actual scene, the corresponding improved algorithm is designed. In the daytime vehicle detection, the traditional lane line detection algorithm based on Hough transform can not extract the curve lane line. In this paper, a new algorithm for lane fitting based on symmetric search is proposed. The fitting error is reduced. A verification algorithm based on PCA dimension reduction algorithm is designed for the low efficiency of the method of directly inputting the suspected vehicle target results into the SVM classifier for classification. That is, the detection efficiency of the algorithm is improved by extracting the eigenvector of the preliminary detection results, which simplifies the input of the classifier. Aiming at the situation that the traditional algorithm can detect the other luminous background by using a single color threshold, the HSV color model of the red halo and the white center of the taillight is established. The two colors are extracted to identify the vehicle taillights, and the evidence fusion algorithm is used to verify the detected taillights on the basis of obtaining the initial detection of the vehicle taillights. The vehicle tracking algorithm based on motion model and tracking queue is designed. The motion model is used to estimate the position of vehicle target in the next frame, and the image matching algorithm is used to search the vehicle target around the predicted position. The motion model is modified with the search results, and the vehicle tracking is accomplished by setting up the tracking queue of the target. The results of processing the video taken on the actual road scene show that, In the environment of Windows 10 operating system and 3.2GHz CPU, the vehicle detection and tracking method designed in this paper can accurately detect forward vehicles and achieve 20 frames per second detection speed, which can meet the real-time requirements.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U463.6;TP277

【参考文献】