全景视觉环境避障测距方法研究

发布时间：2018-08-09 15:45

【摘要】：在移动机器人自主避障导航领域中,视觉传感器在获取周围环境信息方面有着许多优势,例如,图像信息丰富,多个视觉传感器之间联合工作时相互干扰较小。而相比传统视觉的狭窄视场,全景视觉拥有的广阔视野能够弥补传统视觉在视场上的缺陷。因此,全景视觉开始广泛应用于自主机器人导航、三维重构、视频监控等领域。国内外专家学者对全景视觉已经作了许多方面的应用研究,但是还存在许多不足。目前,绝大部分应用于机器人避障方向的全景相机使用的平台都是单视点全景相机,获得的图像往往失真严重。多目全景相机可以同时获得360度的多张图像,并且畸变较小,而针对多目全景视觉的应用极少。因此,寻求一种适用于多目全景相机的稳定、有效、方便的测距方式具有重要的研究意义。同时,本实验室在Bug避障算法已经取得了一定的工作成绩。Bug避障算法是一要求传感器具有360度探测范围的简单避障算法。本实验室基于激光测距仪已经实现了非360度探测范围的Bug避障算法,并且能以平滑的路径绕开障碍物达到终点。但是,由于不能得到全方位的环境信息,机器人在避障过程中需要频繁转向,导致避障效率不高。因此,本文针对上述存在问题,尝试将激光测距仪置换为360度视觉传感器,利用全景相机Ladybug3系统平台提出了一种基于单双目结合的全景测距算法。对如何实现全景避障测距进行了深入研究。为实现该测距算法,本文作了如下工作:(1)研究了全景相机测距的基础知识。主要内容包括相机的标定:对比几种标定法后,决定采用张正友标定法;测距图像预处理:采用直方图均衡化和中值滤波增强图像对比度和去除噪声;立体匹配:分析了几种匹配方法,采用改进的SURF匹配法提取特征点对的坐标,提高了匹配鲁棒性。(2)阐述了单双目融合全景测距原理。第一,探讨了如何确定障碍物是在重叠区域或非重叠区域;第二,确定了在重叠区域采用双目测距,在非重叠区域采用单目测距的测距机制;第三,利用透视原理实现双目测距,对测距原理作了详细推导;第四,采用非线性回归建模方法实现单目测距,并阐述了建立非线性回归模型具体过程。(3)完成了单双目测距实验。选择学校操场作为试验场地,进行了标定实验,测距图像预处理,测距图像立体匹配实验,双目测距试验和单目测距实验。分别对实验结果和误差进行了对比分析。实验结果表明,双目测距结果的测距误差在1.08%~4.48%范围,最大误差4.48%出现在4m处。单目测距结果的误差范围为0.27%~12.57%,最大误差为1.4 m时的12.57%,误差的变化较为随机,并不随着距离的增大而增大,但是误差明显比双目测距要大。上述误差均在可接受范围内,可用来实现避障。
[Abstract]:In the field of autonomous obstacle avoidance navigation of mobile robots, visual sensors have many advantages in obtaining information about the surrounding environment. For example, the image information is abundant and the interference between multiple vision sensors is small when they work together. Compared with the narrow field of view of traditional vision, panoramic vision has a wide field of vision to make up for the shortcomings of traditional vision in the field of view. Therefore, panoramic vision has been widely used in autonomous robot navigation, 3D reconstruction, video surveillance and other fields. Experts and scholars at home and abroad have done a lot of research on panoramic vision, but there are still many shortcomings. At present, most of the panoramic cameras used in obstacle avoidance are single view panoramic cameras, and the images are often distorted seriously. Multi-view panoramic camera can obtain 360-degree images at the same time, and the distortion is small, but there are few applications for multi-view panoramic vision. Therefore, it is of great significance to seek a stable, effective and convenient ranging method for multi-view panoramic cameras. At the same time, our lab has made some achievements in the Bug obstacle avoidance algorithm. The Bug obstacle avoidance algorithm is a simple obstacle avoidance algorithm which requires the sensor to have a 360-degree detection range. Based on the laser rangefinder, a non-360 degree range Bug obstacle avoidance algorithm has been implemented in our laboratory, and a smooth path can be used to bypass the obstacle to reach the end point. However, due to the lack of comprehensive environmental information, the robot needs to turn frequently in the process of obstacle avoidance, which leads to the low efficiency of obstacle avoidance. Therefore, aiming at the above problems, this paper attempts to replace the laser rangefinder with a 360-degree vision sensor, and puts forward a panoramic ranging algorithm based on a panoramic camera Ladybug3 system platform. How to realize panoramic obstacle avoidance ranging is studied. In order to realize the ranging algorithm, the following work is done: (1) the basic knowledge of panoramic camera ranging is studied. The main contents include camera calibration: after comparing several calibration methods, Zhang Zhengyou calibration method is adopted; ranging image preprocessing: histogram equalization and median filter are used to enhance image contrast and remove noise; Stereo matching: several matching methods are analyzed and the coordinates of feature pairs are extracted by improved SURF matching method to improve the matching robustness. (2) the principle of monocular binocular fusion panoramic ranging is expounded. First, it discusses how to determine whether the obstacle is located in an overlapping area or a non-overlapping area; secondly, it determines a binocular ranging mechanism in an overlapped area and a single visual distance measurement mechanism in a non-overlapping region; and third, Using the perspective principle to realize binocular ranging, the principle of ranging is deduced in detail. Fourthly, the single visual distance is realized by nonlinear regression modeling method, and the concrete process of establishing nonlinear regression model is described. (3) Mono-binocular ranging experiment is completed. The school playground is chosen as the test site, and the calibration experiment, ranging image preprocessing, ranging image stereo matching experiment, binocular ranging experiment and single eye distance measurement experiment are carried out. The experimental results and errors are compared and analyzed respectively. The experimental results show that the ranging error of binocular ranging results is in the range of 1.08% and 4.48%, and the maximum error is 4.48% at 4m. The error range of the single visual distance is 0.27 and 12.57, and the maximum error is 12.57 when the maximum error is 1.4 m. The variation of the error is random and does not increase with the increase of distance, but the error is obviously larger than that of binocular ranging. The above errors are within acceptable range and can be used to avoid obstacles.
【学位授予单位】：华南农业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP242

【相似文献】