基于并行计算的苹果采摘机器人关键技术研究

发布时间：2018-10-24 22:16

【摘要】：机器人避障预测模型、图像处理和存储优化是苹果采摘机器人软件系统中的重要研究内容。这些模块的算法优化和并行处理是提升苹果采摘机器人软件系统性能的捷径之一。本文描述了一种可扩展的苹果采摘机器人并行系统的软硬件框架结构,说明在苹果采摘机器人技术中应用并行技术的可行性和必要性。针对苹果采摘机器人上位机软件系统中的相关算法——基于决策树的机器人避障预测模型、含噪苹果图像聚类分割和匹配等模块的设计特点及相关性能等进行了研究分析。以并行化处理为主线,运用MapReduce编程模型和集群等并行技术,设计了这些算法的优化算法或并行算法,以提高相关算法的效率或识别准确率。主要研究工作和结论及创新点如下:(1)通过机器人避障样本建立状态空间映射到决策空间关系矩阵,是决策树判定的变换空间。传统的决策树生成算法对大样本的数据挖掘和处理能力是有限的。本文提出一种基于MapReduce和属性集依赖度的预测模型并行生成算法,可对机器人避障预测中的决策树生成进行并行处理。该算法采用属性集依赖度作为测试属性的选择标准来避免ID3算法不易去除噪声以及忽视属性间相互关联等缺点,在考虑属性集和属性集内元素相互依赖的基础上对属性或属性集进行约简,从而去掉冗余属性(集)。通过机器人避障操作实例的仿真结果验证了基于MapReduce和属性集依赖度并行决策树生成算法可以处理大规模数据样本的分类与决策问题,同时与之前算法进行了复杂度等比较,具有较好的可扩展性和较高的分类效率。(2)基于空间特征的谱聚类含噪苹果图像分割的并行优化算法围绕图像去噪、优化和并行谱聚类等问题而设计。其基本思路是构造图像的三维空间特征点紧致性函数,用以构造邻近点的相似矩阵实现图像的去噪效果;再利用离群点矩阵拆分并由其它剩余列向量线性表示,对相似矩阵进行离群点调优实现聚类优化,降低离群点对谱聚类算法聚类的精确度的影响,最后设计MapReduce函数来并行化处理。为了验证基于空间特征的谱聚类方法的去噪效果以及离群点优化的聚类优化,对两幅苹果图像添加不同程度的高斯和椒盐噪声(方差分别为0.01、0.05和0.1的高斯噪声和概率分别为0.01、0.05和0.1的椒盐噪声),分别求出谱聚类方法、基于空间特征的谱聚类方法、离群点优化方法和基于MapReduce的离群点优化方法的苹果目标图像的分割图,并计算四类方法的分割准确率。结果表明:谱聚类方法受噪声的影响较大;基于空间特征的谱聚类方法的分割效果受噪声的影响较小,但在边界区域仍然有很多错分的像素;离群点优化方法和基于MapReduce的离群点优化方法在边界区域的分割要优于基于空间特征的谱聚类方法;在设定的实验条件下,其分割结果准确率相对于基于空间特征的谱聚类方法和传统的谱聚类方法可分别提高5%~6%和9%~25%,且其后者与前者的时间加速比为11%左右。(3)提出一种基于集群和降低维数的图像匹配并行处理方法,可以处理大数据量的灰度阵列信息,在不降低相似度量的前提下提高其匹配效率。该方法属于快速一维投影模板匹配算法,第一步是降维投影其二维实时图像,其次是差分量化其投影值,形成由0、1逻辑数字组成的描述图像与模板的特征字符串,引入基于同构集群的分级嵌套串匹配并行算法对图像和模板直接进行匹配。验证了这种新的模板快速并行匹配算法的并行图像处理的可行性及分析比较结果。(4)在上述算法实验的过程中,实验平台Hadoop的HDFS是基本并行存储结构,图像等文件会遇到访存不均衡的问题,它有很强的访存偏向性与时效性。本文注意到对其做存储和预取优化能改进和提高系统效率。提出将不同的存储单元(异构)组合成为更多不同层级的存储设备的方案,把本次任务需用的文件存储在更合适的磁盘位置,这样就可以在充分考虑成本因素的条件下搭建高性能存储,并利用预取技术减少MapReduce任务等待时间。综上,通过对几种典型软硬件并行处理技术进行较为深入的研究,将这些技术应用于苹果采摘机器人软件系统中主要算法(包括机器人避障预测与决策、图像处理和并行存储优化)中。通过实例仿真以及平台试验的运行结果分析并行化过程和其中中间代码的生成以及复杂度分析得出了一些有益结论。也可为其它领域相通的算法运用MapReduce、集群等手段进行并行化改进处理提供借鉴和参考。
[Abstract]:Robot obstacle avoidance prediction model, image processing and storage optimization are important research contents in apple picking robot software system. The algorithm optimization and parallel processing of these modules is one of the shortcuts to improve the performance of apple picking robot software system. This paper describes a software and hardware frame structure of an extensible apple picking robot parallel system, and explains the feasibility and necessity of applying parallel technology in apple picking robot technology. Aiming at the related algorithms in the software system of apple picking robot, the design features and related performances of the module based on decision tree-based robot obstacle avoidance prediction model, noise-free apple image clustering and matching are studied and analyzed. The optimization algorithm or parallel algorithm of these algorithms is designed to improve the efficiency or accuracy of the algorithm by parallelizing the main line, using the MapReduce programming model and the cluster and other parallel technologies. The main research work and conclusion and innovation point are as follows: (1) the state space is mapped to the decision space relation matrix through the robot obstacle avoidance sample, which is the transformation space of decision tree decision. Traditional decision tree generation algorithm is limited to data mining and processing capability of large samples. A prediction model parallel generation algorithm based on MapReduce and attribute set kurtosis is presented in this paper. The decision tree generation in robot obstacle avoidance prediction can be processed in parallel. The algorithm avoids the disadvantage that the ID3 algorithm is not easy to remove noise and ignores the correlation between attributes by adopting the attribute set kurtosis as the selection criterion of the test property, reduces the attribute or the attribute set on the basis of taking into account the element interdependence in the attribute set and the attribute set, thereby removing the redundant attributes (set). Based on the simulation results of the robot obstacle avoidance operation example, the classification and decision-making of large-scale data samples can be processed by a parallel decision tree generation algorithm based on MapReduce and attribute set. At the same time, the complexity and the like are compared with the previous algorithm. has better extensibility and higher classification efficiency. (2) The parallel optimization algorithm based on spatial feature-based spectral clustering with noisy apple image segmentation is designed around the problems of image de-noising, optimization and parallel spectral clustering. The basic idea is that the three-dimensional space feature point compactness function of the image is constructed, so that the de-noising effect of the image is realized by constructing the pixel matrix of the adjacent points, the outlier matrix is used for splitting and linear representation by the other residual column vectors, the clustering optimization is realized by the outlier adjustment of the pixel matrix, The influence of outliers on the accuracy of the clustering of spectral clustering algorithms is reduced, and the MapReduce function is designed to parallelize. In order to verify the de-noising effect of spectral clustering method based on spatial features and the optimization of outliers, we add different degrees of Gaussian and salt-pepper noise to two apple images (the Gaussian noise and the probability of 0. 01, 0. 05 and 0. 1, respectively), The spectral clustering method, the spectral clustering method based on the spatial feature, the outlier optimization method and the apple target image segmentation map based on MapReduce are respectively obtained, and the segmentation accuracy of the four methods is calculated. The results show that the spectral clustering method is affected by the noise, and the partition effect of the spectral clustering method based on the spatial features is less affected by the noise, but there are still many wrong pixels in the boundary area. The outlier optimization method and the outlier optimization method based on MapReduce are superior to the spatial feature-based spectral clustering method in the segmentation of the boundary region; under the set experimental conditions, The accuracy of segmentation can be improved by 5% ~ 6% and 9% ~ 25% respectively with respect to the spectral clustering method based on spatial features and the traditional spectral clustering method, and the latter's time-acceleration ratio is about 11%. (3) An image matching parallel processing method based on clustering and reducing dimension number is proposed. The gray array information of large data volume can be processed, and its matching efficiency is improved without reducing similar measures. the method belongs to a fast one-dimensional projection template matching algorithm, A hierarchical nested string matching parallel algorithm based on isomorphic clusters is introduced to directly match the images and templates. This paper verifies the feasibility of parallel image processing of the new template fast parallel matching algorithm and compares the results. (4) In the process of the above algorithm experiment, the HDFS of the experimental platform Hadoop is the basic parallel storage structure, and the files such as the image meet the problem of non-equilibrium in the interview, and it has strong follow-up bias and timeliness. It is noted here that storage and pre-fetch optimization can improve and improve system efficiency. A scheme for combining different storage units (heterogeneous) into storage devices at different levels is proposed, and the files needed by this task are stored in a more suitable disk position, so that high-performance storage can be constructed under the condition that the cost factors are fully considered, and utilizing the pre-fetch technology to reduce the task waiting time of the MapReduce task. In this paper, by comparing several typical software and hardware parallel processing techniques, these techniques are applied to the main algorithms of apple picking robot software system (including robot obstacle avoidance prediction and decision-making, image processing and parallel storage optimization). Some useful conclusions are obtained through the example simulation and the running result analysis parallelization process of the platform test and the generation of intermediate code and the complexity analysis. It can also use MapReduce, cluster and other means to provide reference and reference for parallelization improvement.
【学位授予单位】：江苏大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：S225.93;TP242

【相似文献】