无人驾驶汽车环境信息提取及运动决策方法研究

发布时间：2018-05-11 13:08

本文选题：无人驾驶汽车 + 单目视觉　；参考：《长安大学》2016年博士论文

【摘要】：随着汽车保有量的增长,越来越多的道路交通事故也给社会和人民造成了巨大的损失。其中,汽车驾驶人的危险驾驶行为是导致道路交通事故频频发生的主要原因。无人驾驶汽车因其无需人类驾驶操纵的特点具有广阔的应用前景。在无人驾驶汽车的行驶过程中,如何实时、鲁棒地提取行驶环境信息,以及在获得信息的基础上进行合理的运动决策是实现其安全、高效自主驾驶的关键,也是无人驾驶汽车研究中的难点和热点。论文依托国家自然科学基金重大研究计划项目(90920305)“无人驾驶车辆智能测试环境研究与开发”和中央高校基金创新团队项目(CHD2011TD006)“基于视觉信息的无人驾驶智能车辆关键技术研究”对无人驾驶汽车环境信息提取及运动决策方法展开研究,以实现无人驾驶汽车安全、高效、智能地行驶。本文的研究内容主要包括:(1)视觉图像数据采集模型和预处理研究。以无人驾驶汽车坐标系作为约束条件,建立视觉图像数据采集模型;针对图像采集质量易受行驶环境影响而造成特征难以提取的问题,研究多尺度Retinex图像增强算法和传统中值滤波算法的改进优化算法,并进行静态离线对比试验。(2)针对复杂道路环境下车道标线检测算法鲁棒性较差的问题,提出面向图像像素点的改进道路图像分割方法以深度挖掘车道标线轮廓信息;在此基础上提出基于抽样行双向扫描和成像模型约束候选特征点相结合的车道标线检测优化算法。为了实现车道标线检测与跟踪模块的有效切换,建立置信度判别模块和失效判别模块。(3)针对非结构化道路边界检测效率和鲁棒性之间难以平衡的问题,提出一种基于置信概率的分块分类方法提取道路边界的特征点,在此基础上运用改进的最小二乘法完成非结构化道路模型参数求解,并进行静态离线对比试验。(4)针对无人驾驶汽车对前方车辆识别定位准确性及稳定性要求高的问题,提出一种基于视觉传感器和64线三维激光雷达信息融合的前方车辆识别算法。通过融合64线三维激光雷达提取的障碍物位置信息,确定图像中前方车辆的感兴趣区域;以类Haar-HOG融合特征作为目标车辆描述方法,采用AdaBoost算法离线训练获得的级联分类器进行前方车辆辨识;对因遮挡问题未被识别出前方车辆的感兴趣区域,提出基于激光雷达坐标系下位置关系信息的再确认方法。(5)无人驾驶汽车运动决策建模方法研究。以宏观行驶规划为前提,在环境信息提取的基础上,结合无人驾驶汽车的自身运动状态,对其在微观动态交通环境下的两类基本运动模式进行深入研究,设计无人驾驶汽车运动模式的决策条件及对应目标量;在此基础上建立基于决策树的运动决策模型;最后,通过构建微观动态交通仿真环境对其进行合理性验证。(6)搭建基于上位机组件的无人驾驶汽车平台,并对其广义视觉传感系统参数进行标定,在此基础上进行道路试验,以验证论文提出的环境信息提取方法的有效性和运动决策模型的合理性。
[Abstract]:With the increase of car ownership, more and more road traffic accidents have caused great losses to the society and the people. Among them, the dangerous driving behavior of the motorists is the main cause of the frequent occurrence of road traffic accidents. The unmanned vehicle has a broad application prospect because of its characteristics without human driving. In the course of driving a vehicle, how to extract the information of the driving environment in real time, and to make a reasonable decision on the basis of obtaining information is the key to realizing its safety, high efficiency and autonomous driving. It is also a difficult and hot spot in the research of unmanned vehicle. The article relies on the National Natural Science Foundation of the National Science Foundation. Objective (90920305) "research and development of intelligent testing environment for unmanned vehicles" and the central university fund innovation team project (CHD2011TD006) "Research on the key technology of unmanned intelligent vehicle based on visual information" to study the extraction and decision method of unmanned vehicle environment information and movement, in order to realize the safety of unmanned vehicle. The main contents of this paper are as follows: (1) the model of visual image data acquisition and the research of preprocessing. Using the unmanned vehicle coordinate system as a constraint condition, the visual image data acquisition model is established, and the multi-scale Reti is studied for the problem that the quality of the image acquisition is easily affected by the driving environment and the feature is difficult to extract. NEX image enhancement algorithm and traditional median filter algorithm improved optimization algorithm, and static off-line contrast test. (2) aiming at the problem of poor robustness of lane marking detection algorithm in complex road environment, an improved road image segmentation method oriented to image pixels is proposed to dig the contour information of lane marking in depth. An optimization algorithm for lane marking detection based on combined sampling row bidirectional scanning and imaging model constraint candidate feature points is proposed. In order to realize the effective switching of lane marking detection and tracking module, confidence level discrimination module and failure discrimination module are established. (3) it is difficult to balance the efficiency and robustness of unstructured road boundary detection. In this paper, a block classification method based on confidence probability is proposed to extract the characteristic points of the road boundary. On this basis, the improved least square method is used to solve the parameters of unstructured road model, and the static off-line comparison test is carried out. (4) the requirements for the accuracy and stability of unmanned vehicle for the identification and positioning of the vehicle ahead are high. A new vehicle recognition algorithm based on visual sensor and 64 line 3D laser radar information fusion is proposed. The area of interest in the vehicle in front of the image is determined by integrating the information of the obstacle location extracted by the 64 line 3D laser radar, and the Haar-HOG fusion feature is used as the target vehicle description method and the AdaBoost algorithm is used. The cascade classifier obtained by off-line training is used to identify the vehicle ahead, and a reconfirmation method based on the location relationship information in the laser radar coordinate system is proposed. (5) a research on the modeling method for the decision making of the unmanned vehicle motion. On the basis of this, combined with the self driving state of unmanned vehicle, this paper makes a thorough study of the two basic models of motion in the micro dynamic traffic environment, designs the decision conditions and corresponding targets of the unmanned vehicle movement mode, and builds a decision model based on the decision tree. Finally, the microcosmic model is built. The dynamic traffic simulation environment is proved to be reasonable. (6) to build an unmanned vehicle platform based on the upper computer components, and to calibrate the parameters of its generalized visual sensing system. On this basis, a road test is carried out to verify the validity of the method of extracting environmental information and the rationality of the model of motion decision.

【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：U463.6

【相似文献】