The Key Technology Based On Parallel Computing Application In Apple Harvesting Robot

Robot obstacle avoidance prediction model, image processing and storage optimization are the important research content in the apple harvesting robot software system. Algorithm optimization and parallel processing of these modules can enhance the performance of apple harvesting robot software system. The hardware and software framework of the scalable parallel apple harvesting robot system are present in the paper. The parallel technology has great significance in apple harvesting robot technology. Design features and related properties of the robot obstacle avoidance prediction model, clustering segmentation of noisy image and image matching were studied and analyzed in the related algorithms of the apple harvesting robot software system. By setting parallel processing as the main line, the parallel processing scheme of human-machine interface module’s software system is designed, and the implement methods and techniques of the software parallelization are discussed, improvement of some algorithms or parallel algorithms were designed. To improve the efficiency or recognition accuracy of related algorithms, we research on the merge between the related algorithms of apple harvesting robot software system and the parallel technologies, such as MapReduce and computer cluster, hoping to improve the real-time performance of the software systems. The main research works and innovation points are as follows:(1)Through the robot obstacle avoidance sample, the relationship matrix between the state space and decision space is established. The traditional decision tree algorithms have limited ability to solve the large-scale data mining and processing. In the paper, a parallel decision tree generation algorithm based on MapReduce and dependency degree of attribute set is proposed, so it can make parallel processing for robot obstacle avoidance prediction and decision tree generation. ID3 algorithm is difficult to remove noise and the relationship among the attributes may be ignored, for avoiding these disadvantages, the algorithm uses the dependency degree of attribute set as the selection criteria of test attribute. Based on consideration of the interdependence between attribute sets and elements of the attribute sets, attributes or attribute sets are reduced so as to remove redundant attributes and the specific algorithm of attribute set dependence degree is given. We can see from the simulation results of the machine obstacle avoidance experiment: a decision tree generation parallel algorithm can deal with the classification and decision problem of large scale data samples, meanwhile compared complexity with the traditional algorithm, and it also has a good scalability and high classification efficiency.(2)Based on three-dimensional space feature, the spectral clustering’s parallel optimization algorithm is designed to solve the issues of image de-noising, optimization and parallel spectral clustering. The basic idea of the algorithm is that the similar matrix of the image feature points in three-dimensional space is constructed. The outliers are tuned. The MapReduce parallel programming is used. By using linear representation of outliers in the similarity matrix to tune outliers, the spectral clustering noisy image segmentation can improve. Outliers’ impact on the clustering accuracy of spectral clustering algorithm should be reduced. MapReduce functions are designed to do parallelization processing. At last, we conduct experiment to validate and analyze the algorithm. Different levels of gaussian noise and salt and pepper noise are added to two apple image, and the results show that: the segmentation effect of spectral clustering method is strongly influenced by noise; the segmentation effect of spectral clustering method based on space feature is slightly affected by noise, but there are still a lot of pixel points recognized wrongly in the border region; the segmentation effect of outliers optimization method and an optimization method based on MapReduce are superior to the spectral clustering method based on space feature in the border area; In the condition of setting experiment, the segmentation accuracy compared with the spectral clustering method based on space feature and traditional spectral clustering method can be respectively increased by 5% ~ 6% and 9% ~ 25%, and the time acceleration ratio between the latter and the former is about 11%.(3) In the paper, a new parallel image matching algorithm that is based on cluster and dimension reducing is proposed, and it can process grayscale array information on large scale, and improve the matching efficiency under the premise of not reducing the similarity. This method is a fast template matching algorithm of one-dimensional projection, firstly, the 2D images are projected into one-dimensional images, and the one-dimensional projection values will be quantified differentially, and then get 0, 1 digit character string to describe the image and the template, the parallel string matching algorithm is introduced for the image and the template directly. This algorithm is based on homogeneous cluster and hierarchical nesting. The feasibility of the algorithm is verified.(4) In the course of the above algorithm experiment, HDFS is the parallel storage structure in the experimental platform Hadoop, and files such as images will encounter the in-balanced access problem, which has a strong access bias and timeliness. By storaging and prefetching optimizing, it can improve system efficiency. The scheme of combining different storages unit into more different levels storage devices is proposed, and files needed in the task are stored in a more appropriate disk location, so you can build high-performance storage under the consideration of the cost factor, and pre-fetching techniques are used to reduce the waiting time of the Map Reduce task.Through the deeply research on the parallel processing technologies of several typical hardware and software, we apply these technologies to the apple harvesting robot software system’s main algorithms, which is consisted of robot obstacle avoidance prediction model, image processing and parallel storage optimization. The parallel process and the generation of intermediate code are analyzed through the simulation experiment, and then some useful results are obtained. It also can provide reference for the parallel algorithm in other fields, which also uses the MapReduce and the computer cluster technology.