基于线结构光的水障碍、凹凸障碍物远距离感测技术

发布时间：2018-05-10 22:30

本文选题：无人车 + 全天时检测识别　；参考：《北京理工大学》2016年博士论文

【摘要】：野外环境下,地面水障碍、凹凸障碍对无人车的安全行驶造成了极大的威胁,并且在很大程度上制约了行进的速度。目前复杂野外环境下,采用三维激光雷达进行无人车全天时障碍识别,价钱昂贵、处理速度慢,效果不佳,是国内外关注的一个难题。因此,对无人车行驶过程中地面障碍进行快速、有效识别具有十分重要的现实意义。本文结合当前无人车的应用研究,提出了一种基于线结构光视觉系统的多传感器水障碍和凹凸障碍物的检测方法,实现了对车辆前方最远可到百米的障碍物全天时检测识别。本文主要研究内容和研究结果如下：第一,研制了无人车水障碍和凹凸障碍物检测识别系统。硬件系统主要由线结构光视觉传感器系统、位姿方位系统组成；对线结构光传感器的设计进行了具体计算、对障碍识别原理进行了阐述。软件处理系统对线结构光传感器得到的图像和激光雷达的点云数据进行具体处理和数据融合。经过理论分析计算和实验验证,本文研制完成了能用于野外地面、车辆前方百米内、昼夜使用的水障碍、凹凸障碍物检测识别的线结构光视觉传感器系统。第二,为了实现障碍物识别多传感器的信息融合,对研制的线结构光传感器进行了特性分析和标定,建立了线结构光传感器的完整模型；在分析了现有线结构光常见标定方法的基础上,提出了适合本文传感器系统标定的方法,提出了确定空间点在三维摄像机坐标系下三维坐标的方法,系统提出和推导了不同障碍特征尺寸计算的公式。第三,对线结构光传感器视觉系统的图像识别处理进行了研究,主要包括线结构光光条的特征分析、线结构光光条处理和障碍识别依据等几个方面,最后通过实验进行了算法验证。对于得到的线结构光图像,首先用本文提出的RAMF算法滤除图像中的椒盐噪声；然后提取光条感兴趣区域(ROI)；再在Steger提出的基于Hessian矩阵进行条纹中心提取的基础上,进行Gaussian卷积递归运算中的参数计算,由此计算Hessian矩阵的特征值和特征向量；最后根据计算的光条法线方向,计算光条的中心坐标。对于水障碍,提出了水障碍快速、有效提取的“黑洞”算法。最后,通过实验验证了传感器在不同时段对不同障碍快速识别算法的有效性。第四,对线结构光障碍物检测传感器的信息与激光雷达数据融合处理进行了研究,复杂环境下使用单个传感器难以满足全天时障碍的快速、有效识别。首先,对64线三维激光雷达原理和数据格式进行介绍；借助点云库PCL,采用VS2010+Matlab混合编程的方式编写了点云采集、处理软件；并进行了障碍检测实验。之后,基于D-S证据理论对障碍目标进行识别融合,提出了改进的Dempster合成规则,得到了可以实现障碍物判断的D-S融合输出值和车辆可通行的点图。
[Abstract]:In the field environment, the surface water obstacle and bump obstacle pose a great threat to the safe driving of the unmanned vehicle, and to a great extent restrict the speed of moving. At present, in the complex field environment, it is a difficult problem to use 3D lidar to identify all-day obstacle of unmanned vehicle, which is expensive, slow and ineffective. Therefore, it is of great practical significance to identify the ground obstacles quickly and effectively. In this paper, based on the current research on the application of unmanned vehicle, a multi-sensor detection method of water obstacle and bump obstacle based on linear structured light vision system is proposed, which realizes the detection and recognition of obstacles up to 100 meters in front of the vehicle. The main research contents and results are as follows: firstly, the detection and identification system of water obstacle and bump obstacle for unmanned vehicle is developed. The hardware system is mainly composed of linear structured light vision sensor system, position orientation system, the design of line structured light sensor is calculated in detail, and the principle of obstacle identification is expounded. The software processing system processes and fuses the image and point cloud data from the linear structured light sensor. After theoretical analysis and experimental verification, a linear structured light vision sensor system is developed, which can be used to detect and identify water obstacles and concave obstacles in the field, within 100 meters in front of the vehicle. Secondly, in order to realize the information fusion of obstacle recognition multi-sensor, the characteristic analysis and calibration of the developed linear structured light sensor are carried out, and the complete model of the linear structured light sensor is established. Based on the analysis of the common calibration methods of linear structured light, the calibration method suitable for the sensor system in this paper is put forward, and the method of determining the three-dimensional coordinates of spatial points in the 3D camera coordinate system is presented. The formulas for calculating the characteristic dimensions of different obstacles are put forward and deduced systematically. Thirdly, the image recognition processing of the vision system of the linear structured light sensor is studied, which mainly includes the characteristic analysis of the line structure light strip, the processing of the line structure light strip and the basis of obstacle recognition, and so on. Finally, the algorithm is verified by experiments. For the linear structured light image, we first use the RAMF algorithm proposed in this paper to filter the salt and pepper noise in the image, then extract the region of interest of the light strip, then extract the fringe center based on the Hessian matrix proposed by Steger. The eigenvalues and Eigenvectors of the Hessian matrix are calculated by calculating the parameters of the Gaussian convolution recursion operation, and the center coordinates of the optical strip are calculated according to the normal direction of the calculated optical strip. For water barrier, a fast and effective black hole algorithm is proposed. Finally, the effectiveness of the fast identification algorithm for different obstacles is verified by experiments. Fourthly, the information fusion of linear structured light obstacle detection sensor and lidar data fusion is studied. It is difficult to use a single sensor in complex environment to quickly and effectively identify the all-day obstacle. Firstly, the principle and data format of 64-line 3D lidar are introduced. With the help of point cloud library, the point cloud acquisition and processing software is compiled by VS2010 Matlab hybrid programming, and the obstacle detection experiment is carried out. Then, based on D-S evidence theory, the obstacle target is identified and fused, and an improved Dempster synthesis rule is proposed. The D-S fusion output value and the point diagram of the vehicle can be obtained.
【学位授予单位】：北京理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TJ810.3;TP212

【相似文献】