基于多示例学习的机器人目标跟踪技术研究
发布时间：2019-06-10 12:23
【摘要】：随着人工智能技术的不断发展与应用,如今它已上升到国家战略层面的高度。机器人作为人工智能技术的集成者,在实际用途中正在受到更多研究人员的关注。而移动机器人的自主目标识别与跟踪应用也是其智能化技术方面需要解决的一个核心问题。如何将改进的机器学习算法移植到移动机器人上,并且使其处理跟踪时光线变化、遮挡和复杂背景等问题的方法更具有鲁棒性,这是极富挑战性的一项研究技术。该文在MT-R轮式移动机器人平台上,通过改进的机器学习算法,实现了机器人的自主目标跟踪和对自身运动控制的功能,从而进一步实现轮式移动机器人的智能化应用。本文的主要研究内容可以概括成:首先该文回顾了视觉目标跟踪的研究现状,列举出了不同视觉目标跟踪算法所采用的方法并分析了这些算法的缺点。着重说明并分析了基于多实例学习的跟踪方法和结合协同训练的算法。完成了对本文对采用方法理论基础的铺垫。同时也简要讨论了移动机器人在国内外的研究发展状况。然后重点介绍移动机器人内部目标跟踪算法。基于检测的目标跟踪算法通常依靠分类器来区分目标和背景来达到跟踪的目的,在分类器进行学习的时候会对图像分成样本采样和添加标签两个单独的步骤,但是这样选择的样本是无目的性的,导致分类器的效果不稳定。本文结合主动学习的模型,提出一种新的样本选择的算法,基于多实例学习算法的框架,将主动样本选择的策略加入到样本采样和标签分配之间,这样可以选出有助于分类器学习的样本,然后结合协同训练方法,防止由于误差积累而导致的漂移,进一步提高算法性能。通过在标准视频序列上和其他六种算法进行对比实验,结果表明本文方法在目标遮挡、光线变化等复杂条件下表现良好,具有一定的鲁棒性。最后,在MT-R移动机器人上运行本文提出目标跟踪算法,结合机器人硬件驱动策略,实现了移动机器人的自主目标跟踪。并通过在不同的实际场景下进行实验来验证移动机器人目标跟踪的鲁棒性。通过实验结果可以得到,本文提出的目标跟踪算法可以有效的帮助移动机器人在目标遮挡、光线变化等情况下进行鲁棒的跟踪。
[Abstract]:With the continuous development and application of artificial intelligence technology, it has risen to the national strategic level. Robot, as the integration of artificial intelligence technology, is being paid more and more attention by more and more researchers in practical use. The autonomous target recognition and tracking application of mobile robot is also a core problem to be solved in its intelligent technology. How to migrate the improved machine learning algorithm to mobile robots and make its methods to deal with the problems of light change, occlusion and complex background are more robust, which is a very challenging research technology. In this paper, the autonomous target tracking and motion control of the robot are realized on the MT-R wheeled mobile robot platform through the improved machine learning algorithm, so as to further realize the intelligent application of the wheeled mobile robot. The main research contents of this paper can be summarized as follows: firstly, this paper reviews the research status of visual target tracking, enumerates the methods used in different visual target tracking algorithms, and analyzes the shortcomings of these algorithms. The tracking method based on multi-case learning and the algorithm combined with collaborative training are emphasized and analyzed. The foundation of the method theory in this paper is completed. At the same time, the research and development of mobile robots at home and abroad are briefly discussed. Then the internal target tracking algorithm of mobile robot is introduced in detail. The target tracking algorithm based on detection usually relies on the classifier to distinguish the target from the background to achieve the goal of tracking. When the classifier is learned, the image will be divided into two separate steps: sample sampling and tagging. However, the sample selected in this way is purposeless, which leads to the instability of the effect of the classifier. In this paper, based on the active learning model, a new sample selection algorithm is proposed. Based on the framework of multi-case learning algorithm, the active sample selection strategy is added between sample sampling and label allocation. In this way, the samples which are helpful to the learning of classifiers can be selected, and then the collaborative training method can be combined to prevent the drift caused by error accumulation and further improve the performance of the algorithm. Compared with the other six algorithms on the standard video sequence, the results show that the proposed method has good performance and robustness under the complex conditions of target occlusion, light change and so on. Finally, the target tracking algorithm is proposed on MT-R mobile robot, and the autonomous target tracking of mobile robot is realized by combining the hardware driving strategy of the robot. The robustness of mobile robot target tracking is verified by experiments in different practical scenarios. The experimental results show that the target tracking algorithm proposed in this paper can effectively help the mobile robot to track the target effectively under the condition of target occlusion, light change and so on.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP242
本文编号：2496445
[Abstract]:With the continuous development and application of artificial intelligence technology, it has risen to the national strategic level. Robot, as the integration of artificial intelligence technology, is being paid more and more attention by more and more researchers in practical use. The autonomous target recognition and tracking application of mobile robot is also a core problem to be solved in its intelligent technology. How to migrate the improved machine learning algorithm to mobile robots and make its methods to deal with the problems of light change, occlusion and complex background are more robust, which is a very challenging research technology. In this paper, the autonomous target tracking and motion control of the robot are realized on the MT-R wheeled mobile robot platform through the improved machine learning algorithm, so as to further realize the intelligent application of the wheeled mobile robot. The main research contents of this paper can be summarized as follows: firstly, this paper reviews the research status of visual target tracking, enumerates the methods used in different visual target tracking algorithms, and analyzes the shortcomings of these algorithms. The tracking method based on multi-case learning and the algorithm combined with collaborative training are emphasized and analyzed. The foundation of the method theory in this paper is completed. At the same time, the research and development of mobile robots at home and abroad are briefly discussed. Then the internal target tracking algorithm of mobile robot is introduced in detail. The target tracking algorithm based on detection usually relies on the classifier to distinguish the target from the background to achieve the goal of tracking. When the classifier is learned, the image will be divided into two separate steps: sample sampling and tagging. However, the sample selected in this way is purposeless, which leads to the instability of the effect of the classifier. In this paper, based on the active learning model, a new sample selection algorithm is proposed. Based on the framework of multi-case learning algorithm, the active sample selection strategy is added between sample sampling and label allocation. In this way, the samples which are helpful to the learning of classifiers can be selected, and then the collaborative training method can be combined to prevent the drift caused by error accumulation and further improve the performance of the algorithm. Compared with the other six algorithms on the standard video sequence, the results show that the proposed method has good performance and robustness under the complex conditions of target occlusion, light change and so on. Finally, the target tracking algorithm is proposed on MT-R mobile robot, and the autonomous target tracking of mobile robot is realized by combining the hardware driving strategy of the robot. The robustness of mobile robot target tracking is verified by experiments in different practical scenarios. The experimental results show that the target tracking algorithm proposed in this paper can effectively help the mobile robot to track the target effectively under the condition of target occlusion, light change and so on.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP242
【参考文献】
相关期刊论文 前6条
1 王丽佳;贾松敏;李秀智;王爽;;基于改进在线多示例学习算法的机器人目标跟踪[J];自动化学报;2014年12期
2 陈思;苏松志;李绍滋;吕艳萍;曹冬林;;基于在线半监督boosting的协同训练目标跟踪算法[J];电子与信息学报;2014年04期
3 谭民;王硕;;机器人技术研究进展[J];自动化学报;2013年07期
4 王素玉;沈兰荪;;智能视觉监控技术研究进展[J];中国图象图形学报;2007年09期
5 胡斌;何克忠;;计算机视觉在室外移动机器人中的应用[J];自动化学报;2006年05期
6 侯志强;韩崇昭;;视觉跟踪技术综述[J];自动化学报;2006年04期
相关博士学位论文 前2条
1 武玉伟;视觉目标跟踪中的表观建模研究[D];北京理工大学;2014年
2 邱雪娜;基于视觉的运动目标跟踪算法及其在移动机器人中的应用[D];华东理工大学;2011年
相关硕士学位论文 前6条
1 韩信;基于双目视觉的轮式机器人动态避障研究[D];浙江大学;2016年
2 姜卫琳;智能移动机器人均值漂移目标跟踪算法研究[D];西北大学;2014年
3 张全喜;单目视觉自主移动机器人的目标跟踪及误差分析[D];南京理工大学;2014年
4 周秋红;基于多示例学习的运动目标跟踪算法研究[D];大连理工大学;2010年
5 耿盛涛;基于双目视觉的机器人目标检测与跟踪研究[D];江南大学;2010年
6 张亚楠;复杂背景下目标跟踪算法研究及其在移动机器人中的应用[D];华南理工大学;2010年
，本文编号：2496445
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2496445.html
