CASIA OpenIR  > 毕业生  > 博士学位论文
基于球拍运动轨迹的打乒乓球机器人回球策略研究
张鵾1,2
学位类型工学博士
导师谭民
2018-05-28
学位授予单位中国科学院研究生院
学位授予地点北京
关键词打乒乓球机器人 球拍位姿 碰撞反弹模型 对手行为 智能回球
摘要
在机器人领域中,打乒乓球机器人是一个集成度很高的研究平台,其涉及到计算机视觉、智能控制以及环境交互等众多问题。本文在已有的打乒乓球机器人的基础上,重点从对手的击球行为方面进行研究,主要内容如下:
    一、提出了一种基于自适应阈值和特征点统计信息的球拍姿态快速检测方法,克服了环境光照复杂、球拍姿态变化多样等因素带来的干扰问题。将图像的RGB和HSV双颜色空间的特征信息进行融合,实现了特征点检测阈值的自适应调整,提高了算法的鲁棒性。为了快速、准确地获取球拍的位姿,提出了一种基于特征点运动连续性和局部信息统计的方法。
    二、提出了一种对手击球碰撞点状态实时估计的方法。首先,依据乒乓球的飞行物理模型,分别采用正向、逆向预测方法,得到乒乓球的来球和回球轨迹。然后,为准确获取乒乓球拍的实时位姿,提出了一种基于双扩展卡尔曼滤波算法的球拍运动轨迹计算方法。在此基础上,结合乒乓球的来球、回球轨迹以及对手球拍的运动轨迹,设计了一种基于时间和空间双尺度的碰撞点状态估计方法,实现了对对手击球时碰撞点状态的实时估计。
    三、将分布式思想引入乒乓球碰撞模型的建立中,设计了一种基于神经网络群的球拍反弹模型的构建方法。通过对经验数据的分组,建立了一个与数据组对应的神经网络群组,并针对不同的经验数据组,单独训练各个神经网络,规避了错误数据对神经网络训练的影响。同时,离线更新神经网络模型,进一
步提高模型的精准度。
    四、设计了一套基于ROS的击球动作捕捉系统,实时获取对手的击球信息;并研制了一款具有5自由度的轻型仿人机械臂,增强了打乒乓球机器人的灵活性。提出了一套基于对手击球动作的回球策略,通过对球拍运动轨迹的分类和检测,提前预测乒乓球的回球轨迹和击打点位置,加快了乒乓球机器人的反应速度。依靠回球时对手球拍的位置设计了竞技和陪练两种回球模式,提高了打乒乓球机器人的智能性。
    最后,总结本文所取得的研究成果,并对下一步研究提出展望。
其他摘要
In the field of robotics, the table tennis robot is a highly integrated research platform, which involves many issues such as computer vision, intelligent control, environmental interaction and so on. Based on the existing table tennis robot, this thesis focuses on the study of the stroke behavior of opponents. The main contents of this thesis are as follows:
  Firstly, an approach of fast detecting the racket pose is proposed based on a self-adaptive threshold selection scheme and the statistic information of feature points, which overcomes the influence caused by the complicated lighting conditions, the disorderly racket pose and other problems. By fusing the information from the RGB and HSV color spaces, the thresholds for detecting the feature points is determined by itself, and the robustness of the approach is improved. In order to obtain the racket pose quickly and accurately, a method based on the motion continuity and the local information statistics of feature points is proposed.
  Secondly, a method of real-time estimation of the collision status of opponents batting is proposed. At first, the forward and backward trajectories of ping-pong could be predicted based on the physical model of the flight of the ping-pong ball. Then, in order to obtain the real-time racket pose accurately, an approach for calculating the racket trajectory relying on a double extended Kalman filter is presented. On this basis, combined with the ball trajectories and the racket trajectory, a rule to estimate the state of hitting point in real time is designed based on the time and space when the opponent hits the ball.
  Thirdly, bringing the distributed concept into the establishment of the ball rebound model, a method to build a rebound model between the racket and the ball based on neural networks is designed. By grouping the empirical data, a neural network group corresponding to the empirical data set is established. And to avoid that the wrong data influence the training of the neural network, each neural network is trained separately according to different empirical data sets.At the same time, the neural networks are updated off-line to further improve the accuracy of the model.
  Fourthly, a ROS-based system to capture the hitting motion is designed for obtaining the information of the opponent's hitting process in real time. And a lightweight artificial manipulator with 5 degrees of freedom is developed to enhance the flexibility of the table tennis robot. Based on opponent's hitting action, a strategy of returning the ball for the table tennis robot is proposed. Through the classification and detection of the trajectory of the opponent's racket, the trajectory and hitting position of the ball can be predicted in advance, and it also improves the responsiveness of the table tennis robot. To improve the intelligence of the table tennis robot, two modes (competing and sparring) for returning the ball back are designed relying on the position of the opponent's racket when the table tennis robot returns the ball.
  Finally, summarization of the research is addressed with discussions and the prospect of the future research.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20971
专题毕业生_博士学位论文
作者单位1.中国科学院大学
2.中国科学院自动化所
推荐引用方式
GB/T 7714
张鵾. 基于球拍运动轨迹的打乒乓球机器人回球策略研究[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于球拍运动轨迹的打乒乓球机器人回球策略(37196KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[张鵾]的文章
百度学术
百度学术中相似的文章
[张鵾]的文章
必应学术
必应学术中相似的文章
[张鵾]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。