基于单目视觉的车辆检测与跟踪算法研究

发布时间：2018-06-16 07:27

本文选题：单目视觉 + 车道线检测　；参考：《哈尔滨工程大学》2014年硕士论文

【摘要】：随着社会的不断进步,汽车正为越来越多的人所使用,而相应的,交通事故也越来越多。为解决这一问题,越来越多国家开始研究智能交通系统。而智能交通的核心基础就是要检测和跟踪道路上的车辆,根据车辆位置信息来避免交通事故的发生。本文正是采用基于视觉的方法检测和跟踪前方车辆的。车辆检测通常分为两个步骤。首先确定车辆可能存在的区域,这其中包括路旁树木等留下的虚假车辆阴影。接下来就要剔除虚假的车辆阴影,确定车辆的具体位置。正常情况下最有可能会与本车发生碰撞的前方车辆在本车道内,所以本文首先检测车道线,根据车道线缩小车辆检测的范围,提高检测效率和精度。在提取车道线的基础上,利用车辆底部在道路上的阴影与路面灰度值的对比度较大,确定车辆可能存在的区域。再融合图像熵等纹理特征剔除虚假的车辆阴影,准确检测出前方车辆。本文的主要工作如下:1.改进了基于OTSU大津阈值法的自适应二值化方法,采用通过统计道路样本区域灰度值特性的方式来估算道路区域灰度值,这样可以避免传统OTSU对整幅图像统计灰度值时计算量大且有非路面区域干扰的缺点,提高算法实时性和准确性。2.在车道线检测算法中,运用了形态学方法和边缘提取方法后,设计了搜索车道线内侧边缘的扫描算法,并通过对比霍夫变换的算法性能,采用了最小二乘法的拟合车道线方法。为进一步提高算法效率,本文采用了车道线跟踪算法,在前一帧图像的车道线位置左右各扩展50像素范围内搜索,大大降低了车道线检测算法时间。根据检测到的车道线结果,本文计算了每帧图像车辆的偏航角,当偏航角超过给定阈值时即表明车辆即将偏离本车道,此时可发出光声等信号提醒司机采取措施。3.在车辆检测与跟踪算法中,本文在基于阴影检测的算法基础上,结合图像熵值和灰度图像对称性排除虚假车辆区域,检测出车辆在图像中位置信息,并采用基于卡尔曼滤波的跟踪方法,在保证检测精度的同时提高了检测效率,增强了算法的实时性。4.本文建立了安全车距的防碰撞模型,即相对车速与最大制动距离之间的关系,并且给出了基于视觉的测距模型,根据图像中检测到的车辆坐标即可计算出车距,进而估算出碰撞时间。本文使用C++语言利用视觉处理库OpenCV1.0编写了前方车辆检测系统软件,并采集了多段道路视频进行实验。实验结果表明本文算法满足实时性要求,在光照条件良好路段能稳定的跟踪前方车辆目标,对于路况复杂情况也具有一定鲁棒性。
[Abstract]:With the development of society, more and more people are using cars, and accordingly, more and more traffic accidents. In order to solve this problem, more and more countries begin to study its. The core of Intelligent Transportation (its) is to detect and track vehicles on the road and to avoid traffic accidents according to the information of vehicle location. In this paper, vision-based methods are used to detect and track forward vehicles. Vehicle testing is usually divided into two steps. First, identify areas where the vehicle may exist, including false vehicle shadows left by roadside trees and so on. The next step is to remove the false shadow of the vehicle and determine the exact location of the vehicle. Under normal circumstances, the front vehicle most likely to collide with the vehicle is in the driveway, so this paper first detects the lane line, narrows the range of vehicle detection according to the lane line, and improves the detection efficiency and accuracy. Based on the extraction of the lane line, the contrast between the shadow at the bottom of the vehicle on the road and the gray value of the road surface is great, and the possible area of the vehicle is determined. Then fusion the image entropy and other texture features to eliminate the false shadow of the vehicle and accurately detect the vehicle ahead. The main work of this paper is as follows: 1. An adaptive binarization method based on Otsu Otsu threshold method is improved to estimate the gray value of road area by statistics of the gray value characteristics of road sample area. In this way, the traditional OTSU can avoid the disadvantages of large computation and non-road area interference when the whole image is calculated by using OTSU, and improve the real-time and accuracy of the algorithm. In the lane line detection algorithm, after using morphological method and edge extraction method, a scanning algorithm is designed to search the inner edge of lane line. By comparing the performance of Hough transform, the least square method is used to fit the lane line. In order to further improve the efficiency of the algorithm, a lane tracking algorithm is adopted in this paper, which searches within the range of 50 pixels about the location of the lane line of the previous frame image, which greatly reduces the time of the lane line detection algorithm. Based on the detected lane line results, this paper calculates the yaw angle of the vehicle in each frame image. When the yaw angle exceeds the given threshold, it indicates that the vehicle is about to deviate from the driveway. At this time, the driver can be warned to take action by means of light and sound signals. In the vehicle detection and tracking algorithm, based on the shadow detection algorithm, combined with the image entropy and gray image symmetry to eliminate the false vehicle area, the vehicle position information in the image is detected. The tracking method based on Kalman filter is used to ensure the detection accuracy and improve the detection efficiency, and enhance the real-time performance of the algorithm. In this paper, the anti-collision model of safe vehicle distance is established, that is, the relation between relative speed and maximum braking distance, and the distance measurement model based on vision is given. The distance can be calculated according to the vehicle coordinates detected in the image. Then the collision time is estimated. In this paper, we use C language and OpenCV1.0 to compile the software of the vehicle detection system in front, and collect the video of many sections of the road to carry on the experiment. The experimental results show that the proposed algorithm can meet the real-time requirements and can track the vehicle targets stably in good lighting conditions. It is also robust to complex road conditions.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】