【作者】 徐颖；

【导师】 房斌；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 近年来,随着社会经济水平的飞速发展,全国人均汽车持有量快速提升,但交通堵塞和安全等问题也随之而来。为了解决这些问题,各国政府都投入大量经费对智能交通系统其进行研究。智能交通系统中的重要组成部分——智能车是一个集环境感知、决策规划以及辅助驾驶等功能于一体的智能平台。智能车中环境感知系统的关键组成部分是基于视觉的车道线检测。车道线检测技术主要通过分析车载视觉系统中采集的道路图像,实时提取出车道线的相关位置信息,以提供给智能车的决策规划、车道偏离预警等系统使用。因此,如何在复杂道路场景下快速、准确地检测出车道线成为了各科研机构的研究重点。基于纹理和Hough变换的道路检测,因其在复杂道路环境下表现出较强的鲁棒性,成为了研究车道线检测的方法之一。然而,在传统的串行车道线检测算法中,往往存在着数据量大、处理速度较慢、计算复杂度高,耗时显著等问题,难以满足实时性要求。虽然有学者通过引入CPU多核技术来提高Hough变换的计算过程,但加速效果并不明显。图形处理单元(Graphic Process Unit,GPU)因为具有强大计算能力而被广泛用于并行计算。根据GPU面向通用计算领域处理数据的特点与优势,结合图像中值滤波、差分激励和Hough变换的适宜于并行的特点,本文提出了一种基于GPU加速的并行车道检测算法,该方法将车道线检测过程中几种计算复杂度比较高的算法进行并行设计并在计算机统一设备架构(Compute Unified Device Architecture,CUDA)并行编程架构上进行实现,通过加速这些算法,使系统能够更好地满足实时性要求。在公开数据库上的仿真实验结果表明:与其他典型的串行车道线检测算法相比,本文所提算法能同时提高车道线的检测精度和检测速度,具有较好的鲁棒性和实时性。本论文的主要研究内容及贡献包括以下三方面:(1)提出了基于GPU的中值滤波算法并行化方法。在图像预处理时,对原始图像中的感兴趣区域转化成灰度图后,由于噪音的影响,需要使用中值滤波对灰度图进行去噪声处理。然而,中值滤波具有较大的计算量,本文结合CUDA并行编程模型的特点,对原始的快速中值滤波进行并行设计与优化,并在CUDA平台上实现了并行的快速中值滤波算法。与传统的中值滤波方法相比,所提算法在速度上有较大提升。(2)提出了基于GPU的差分激励算法并行化方法。对图像进行去噪声处理后,利用差分激励方法来增强灰度图像中的纹理信息,以便于车道线特征点的提取。然后根据韦伯定理,将能够引起人类视觉系统注意的信息部分提取出来,将灰度图像二值化。在差分激励方法中计算邻域差值时,可以使用多线程同时处理不同的像素点,所以本文对差分激励方法进行并行化处理。通过对并行差分激励算法的设计,所提算法有效地减轻了特征点提取的计算量,提高了算法的实时性。(3)提出了基于GPU的Hough变换算法并行化方法。针对传统的Hough变换检测图像的算法存在着耗费内存空间较大及运算时间较长的缺点,提出了基于GPU的Hough变换算法并行化方法。该方法根据传统Hough变换算法适于并行的特点,设计了相关的并行策略,并在CUDA并行编程模型中进行了实现。此外,本文还对传统的算法进行了优化,采用计算局部最值的方法替代传统的分块求最值,从而提高Hough变换的处理速度和直线的检测精度。更多还原

【Abstract】 In the wake of the rapid development of society,more and more people have their own car in China.However,many troublesome questions are accompanying,such as traffic jams and traffic safety.The intelligent transportation system which aims to solve those traffic problems gets more attention now,and governments have invested a lot of resources to study it in the world.As a key part of intelligent transportation system,the intelligent vehicle is an intelligent platform which integrates multiple functions,such as environment perception,decision-making,and driver assistance,as the key component of environment perception system of intelligent vehicle,the lane detection technology based on computer vision has been the focus of research topics for scholars.The technology extracts location information of the lane in time through analyzing road images captured by vehicle-mounted cameras.The information is used in decision-making system,lane departure warning system,etc.Therefore,how to rapidly and accurately detect lane markings under complex road conditions is an import technology for intelligent vehicle.Because of high robustness in many complex road environment,the road detection algorithm based on texture feature and Hough transform,has become one of the methods to research lane detection.However,there are still some problems about a large amount of data,slow processing speed,high computational complexity and time-consuming in the traditional serial lane detection algorithm.Therefore,it is very difficult to meet the real-time requirements.Although some scholars have adopted a multi-core technology based on CPU to improve the speed of detecting line for Hough transform,the acceleration effect is not obvious.Graphic Processing Unit(GPU)with the powerful computing ability is widely used for parallel computing.According to the characteristics and advantages of GPU in data processing for general purpose computing,combined with the characteristics of image median filtering,differential excitation and Hough transform are suitable for parallelism,this paper proposed a parallel lane detection algorithm based on GPU acceleration.The proposed algorithm implements fast median filtering,differential excitation and Hough transform on compute unified device architecture(CUDA).This algorithm took the advantages of GPU in parallel computation,memory management and reasonably allocated the computational resources and the corresponding computational tasks to the host and device in the lane detection.Experimental results on an open database indicate that the proposed method outperforms the classical approaches,compared with common methods,the proposed method has some improvement in accuracy and speed.The main research content and contribution of this paper includes the following three aspects:Firstly,this paper proposed a special fast median filtering algorithm based on CUDA architecture.In the image preprocessing,after the region of interest in the original image is transformed into a grayscale image,the median filter is used to eliminate the influence of noise on grayscale.However,the median filter has a large amount of computation.In this paper,the characteristics of CUDA parallel programming model are used to optimize the original fast median filter,and parallel algorithm of fast median filter is implemented on CUDA platform.Compared with common methods,the proposed method has some improvement in speed.Then,a GPU-based parallelization method of differential excitation algorithm is proposed.After noise reduction,the differential excitation method is used to enhance the texture information in the gray image to extract the feature points of the lane line.Then according to Weber’s theorem,we can selects the largest part of the information which can cause human visual system’s attention,and the useful information to gain the binary image.According to the characteristics that differential excitation is suitable for parallel computing,we can use multi-threading method to process different pixels when calculate the neighborhood difference in differential excitation method,the proposed algorithm can effectively reduce the computational complexity of feature point extraction and improve the real-time performance of the algorithm by such parallel design.Finally,in view of shortages of large memory space and long computation time of the traditional Hough transform,a parallel Hough transform algorithm based on GPU is proposed in this paper.This algorithm designs the related parallel strategy based on the characteristics of the traditional Hough transform algorithm and implements it in the CUDA parallel programming model.In addition,we do some performance optimization for this algorithm,for example,calculating the part maximum value method instead of the traditional block seeking.The experimental results show that,compare to the typical methods,the proposed algorithm has better accuracy and less computation time.更多还原