基于CUDA的实时目标识别系统的设计与实现

发布时间：2019-04-01 15:08

【摘要】：机器人视觉伺服是一个复杂的系统,能够应用在不同领域。本文以视觉伺服为背景,重点研究基于CUDA的快速目标识别内容。采集相机信息,最终给出识别目标的位置偏差,输出给视觉伺服系统以供控制应用。文中首先重点讨论了目标识别算法的相关问题,包括跟踪方法,识别方法等等。其次重点关注并行优化问题,对实际选择的SIFT,CAMSHIFT等基本模块的算法实现进行了针对CUDA平台的应用优化。最后通过搭载机器人视觉伺服系统,验证了方法的可行性和方法的性能。本文研究内容包括四个部分,各部分内容概括如下:第一部分:重点阐述了论文中所涉及到的几个基本算法的相关内容,包括一些原理的解释,理解和说明。首先简要介绍了项目背景下的算法处理过程,阐述明白实际的输入输出。其次介绍了SIFT的相关基础知识,包括尺度空间的构建,极值点检测,特征点梯度计算,特征描述子计算,特征匹配等。再次介绍了CAMSHIFT的相关基础知识,包括直方图的生成,反向概率投影,图像矩计算,直方图相交等。最后介绍了并行优化的相关内容,包括并行规约,Amdahld定理,Gustafson定理并行优化原理上的东西。第二部分:重点阐述了快速目标识别算法的具体设计。包括具体的应用实现,具体的配合方式等。首先介绍了基于SIFT特征匹配实现稳定的特征匹配,用来提供稳定的特征参考。其次介绍了基于CAMSHIFT跟踪的快速目标ROI获取。最后介绍了算法的评价机制以及识别策略。第三部分:重点阐述了快速目标识别算法的实际并行优化设计。从并行优化的角度上,具体应用相关原理。首先是基于SIFT相关模块进行实际的CUDA框架设计实现。其次是基于CAMSHIFT相关子模块进行实际的CUDA框架设计实现。第四部分:进行了具体的实验。分别从单模块的识别效果,整体的识别效果等,具体展示该方法的实际效果。并且对结果进行了简单的分析和介绍。
[Abstract]:Robot visual servo is a complex system which can be applied in different fields. Under the background of visual servo, this paper focuses on the fast target recognition based on CUDA. The camera information is collected. Finally, the position deviation of the target is identified and output to the visual servo system for control application. In this paper, we focus on the related problems of target recognition algorithms, including tracking methods, recognition methods and so on. Secondly, the parallel optimization problem is focused on, and the algorithm implementation of the basic modules such as SIFT,CAMSHIFT is optimized according to the CUDA platform. Finally, the feasibility and performance of the method are verified by the robot visual servo system. The research contents of this paper include four parts. The contents of each part are summarized as follows: the first part focuses on the related contents of several basic algorithms involved in the paper, including the explanation, understanding and explanation of some principles. Firstly, the algorithm processing process under the project background is briefly introduced, and the actual input and output are explained. Secondly, the basic knowledge of SIFT is introduced, including the construction of scale space, detection of extreme points, gradient calculation of feature points, computation of feature descriptors, feature matching and so on. Thirdly, the basic knowledge of CAMSHIFT is introduced, including histogram generation, backward probability projection, image moment calculation, histogram intersection and so on. Finally, the related contents of parallel optimization are introduced, including parallel specification, Amdahld theorem and Gustafson theorem. The second part focuses on the design of fast target recognition algorithm. Including the concrete application realization, the concrete coordination way and so on. First, the stable feature matching based on SIFT feature matching is introduced, which is used to provide stable feature reference. Secondly, the fast target ROI acquisition based on CAMSHIFT tracking is introduced. Finally, the evaluation mechanism and identification strategy of the algorithm are introduced. The third part focuses on the practical parallel optimization design of fast target recognition algorithm. From the point of view of parallel optimization, the related principle is applied in detail. First of all, the design and implementation of the actual CUDA framework based on the relevant modules of SIFT. Secondly, the design and implementation of the actual CUDA framework based on the CAMSHIFT related sub-module is carried out. The fourth part: the concrete experiment has been carried out. From the single module recognition effect, the overall recognition effect and so on, the actual effect of this method is shown in detail. The results are simply analyzed and introduced.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】