葡萄园植保机器人路径规划算法

为提高植保机器人葡萄园作业在垄行识别和路径规划中的准确度和可靠性,该文提出了一种基于支持向量机(support vector machine,SVM)的多支持向量配比权重进行葡萄园垄行识别与农业机器人作业路径规划的算法。该算法利用Kalman滤波器对由激光雷达扫描获取的粗大实况果园数据信息进行预处理,校正数据中的噪声离群点,然后结合SVM,获得垄行环境中的分割超平面和分类边际线,最后根据样本点与分类边际线存在的几何间隔关系判别各点所占相对权重,获取垄线安全预估测位置并进行农业机器人作业导航线的规划拟合。通过对多个实际样本的试验与测试,拟合导航线与实际垄行中心线平均角度偏差为0.72°,相对植保机器人的平均距离偏差为4.22 mm。试验结果表明,该算法能够有效的识别与定位植保机器人所需导航线的位置,拟合的导航线满足葡萄园植保机器人准确作业的要求。

Path planning algorithm for plant protection robots in vineyard

To meet the requirements of accuracy and reliability of plant protection robot in ridge identification and route planning, also improve the working conditions of farmers, and achieve an unmanned operation purpose, an algorithm based on multi-support-vector proportioning weight of SVM (support vector machine) to identify the ridge line of vineyards, and the path planning of plant protection robots were proposed. The strategy first uses Kalman filter to pre-process coarse orchard data information obtained by Lidar scanning. According to the principle that Kalman filter complies with Gaussian distribution, the prior point between 2 adjacent points was taken as the prior state, and the latter point was used as the observed point to obtain posteriori state estimation, so as to realize data integration and analysis. With its good system state estimation characteristics, the collected data can be used to judge the trend of the ridge line, so as to correct the noise outliers in the data and improve the readability of the data. Then according to the characteristics of the vineyard branch ridge and the characteristics of ridges line with clearly separable spacing, and corresponding to the situation that the ridge line on both sides can be completely separated, the method was combined with SVM linear classification. With the unique advantages of the classification and due to that SVM can search the unique segmentation hyperplane, the maximum interval and segmentation hyperplane, classification margins in the ridge environment could be gotten. The split hyperplane obtained at this time would be between ridge lines. However, there was a big deviation from the angle of the direction of the ridge line and the horizontal distance. It could not meet the precise operation requirements of plant protection robots. It needed further data processing and analysis. In order to obtain accurate position of the center of ridge line, finally, the relative weights were assigned to the sample points of each ridge based on the geometric interval relationship between the sample points on both sides of the ridge and the corresponding SVM classification marginal line. The classification marginal line was reformed according to the number of sample points and the relative weights. According to the condition of the product of the interval relationship between each sample point and the classification margin, their relative weight must be consistent with the quality value of classification margin. The random sampling consistency iteration method (RANSAC) would avoid the error of cost estimate, and could estimate the parameters of the mathematical model from a group of observed data with outliers, so as to obtain the predicted safety location of the ridge. Although the pre-estimated security location of ridge line was not necessarily consistent with the actual location of the vineyard ridge, navigation line could be obtained indirectly by the security ridge line on both sides and the principle of angle bisector which could meet the requirements of precision operation of plant protection robot. Operation guidance line for plant protection robot could be acquired. After a number of actual samples were tested, the average angular deviation between the fitted navigation line and the actual ridge centerline was 0.72°, and the average distance deviation of the relative plant protection robot was 4.22 mm. Experimental results showed that this algorithm could effectively identify and locate the navigation route needed by the plant protection robot. The fitted navigation line could meet the requirements of accurate operation of the plant protection robot in the vineyard. However, the redundancy of the algorithm was relatively large, and the time required to process data in a single time was about 2.05 s. With the accelerated calculation speed of the processor in the future, the algorithm provided in the article can provide a reference solution for such a problem.