果树自动喷药中的双目视觉方法研究

发布时间：2018-02-26 22:43

本文关键词： 自动喷药 SIFT 双目测距视觉SLAM　出处：《烟台大学》2017年硕士论文　论文类型：学位论文

【摘要】：果树自动农药喷洒即移动机器根据周围环境自主完成对农作物的药物喷洒过程,从而代替果农的操作,降低人的劳动强度。此过程可以分为四部分:动力驱动部分、运动控制部分、周围环境识别与决策部分以及喷药执行部分。而周围环境识别与决策可以归结为一个自主导航的问题。同时定位与地图创建(SLAM)问题是自主导航的基础与核心,为自主导航提供了理论依据与解决方案。SLAM即运用相应算法处理传感器获取的信息,完成自身的状态位置估计以及整个任务目标区的环境地图建立,同时利用已知的地图反馈估计自身的位置,从而适应复杂的环境。本文针对未知环境下双目视觉SLAM系统进行研究。对果树的喷药操作,首先需要确定目标物的方位、距离信息。在图像处理领域中,双目测距的方法能够较好的完成对目标物距离的测量,而要获得较高的距离测量精度,就要通过相机标定获取相机焦距参数参与计算。实现双目测距的另一个关键在于选取目标物上一个较为稳定的点,由于SIFT特征点的尺度不变性、旋转不变性以及稳定精确性等特点,本文采用SIFT算法完成目标物稳定点的提取。但接下来的问题是并非所有的SIFT特征点都在植株上,所以应该限定特征点的选取范围。本文在阈值分割算法的基础上进行改进,提出了一种基于最大分量分析的阈值分割算法,这样便可保证所有的特征点都在植株上,另外也减少了特征点数量,提高了算法执行效率。双目测距还需要给出左、右目相机中同一个测量点的图像位置偏移量,而两幅图像中同一特征点是可通过特征点匹配进行识别。针对存在较多错误匹配点的情况,本文对图像中的极限几何约束关系进行改进,提出了一种平行约束关系,结合RANSAC算法较好的实现了错误匹配点的剔除。最后选取中心匹配点对进行距离测量。接下来以此特征测量点的距离信息作为SLAM系统的输入,利用卡尔曼滤波方法进行相关的状态估计矫正以及地图建立进行研究,为后续的相关研究奠定了基础。在实验验证方面,进行了相关背景图下的特征点提取与匹配实验以及双目测距实验,取得了较好的效果。
[Abstract]:The automatic pesticide spraying of fruit trees means that the moving machine completes the spraying process of crops according to the surrounding environment, which replaces the operation of the fruit farmers and reduces the labor intensity of the people. The process can be divided into four parts: the power driving part. The motion control part, the surrounding environment identification and decision making part and the spray execution part. And the surrounding environment recognition and decision can be summed up as a problem of autonomous navigation. At the same time, the problem of location and map creation is the foundation and core of autonomous navigation. It provides the theoretical basis and solution for autonomous navigation. SLAM uses the corresponding algorithm to process the information obtained by the sensor, and completes the estimation of its own state position and the establishment of the environmental map of the whole task target area. At the same time, the known map feedback is used to estimate its position to adapt to the complex environment. In this paper, the binocular vision SLAM system in unknown environment is studied. Range information. In the field of image processing, the binocular ranging method can achieve the target distance measurement better, but to obtain a higher range measurement accuracy, The focal length parameters of the camera must be obtained by camera calibration to participate in the calculation. Another key to achieve binocular ranging is to select a relatively stable point on the target, because of the scale invariance of the SIFT feature points. In this paper, SIFT algorithm is used to extract the stable point of the object, but the next problem is that not all the SIFT feature points are on the plant. In this paper, we improve the threshold segmentation algorithm and propose a threshold segmentation algorithm based on maximum component analysis, which can ensure that all the feature points are on the plant. In addition, the number of feature points is reduced, and the efficiency of the algorithm is improved. The same feature point in two images can be identified by matching the feature points. In view of the fact that there are more mismatched points, this paper improves the limit geometric constraint relation in the image, and proposes a parallel constraint relation. Combined with the RANSAC algorithm, the error-matching points are eliminated. Finally, the center matching points are selected to measure the distance. Then, the distance information of the feature measurement points is used as the input of the SLAM system. The Kalman filtering method is used to study the state estimation correction and map establishment, which lays a foundation for further research. The feature points extraction and matching experiments and binocular ranging experiments under the relevant background images are carried out, and good results are obtained.
【学位授予单位】：烟台大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S49;TP391.41

【参考文献】