基于机器视觉的羽毛球路径跟踪算法的研究及实现

发布时间：2019-03-20 09:54

【摘要】：羽毛球运动是一个实时性要求较高的竞技体育项目,要实现羽毛球机器人的人机对打,需要对高速运动的羽毛球进行实时动态跟踪。动态目标的跟踪问题广泛存在于工业生产、生活以及军事领域,如导弹拦截系统,道路交通实时监控系统。本课题利用双目视觉系统实时捕获动态飞行过程中的羽毛球,并根据羽毛球飞行特性,建立羽毛球空气动力学模型。本课题研究内容如下:(1)分析比较了三种在常用的动态目标捕捉算法:背景差法、光流法和帧间差分法。分析比较这三种算法的优点与缺点。结合羽毛球的飞行特性,建立来基于三帧帧间差分法与光流算法结合的羽毛球动态捕捉算法,并分析其优缺点。(2)建立羽毛球的空气动力学模型。分析羽毛球在空中飞行所受到空气动力和重力对它造成的影响。利用实验分析影响羽毛球飞行轨迹的主要因素。通过高速相机拍摄到的羽毛球飞行轨迹,建立羽毛球稳态飞行模型以及自旋模型。(3)在羽毛球轨迹预测过程中,因为存在着模型误差以及测量误差,通过贝叶斯滤波计算羽毛球状态的最大似然估计。卡尔曼滤波器是线性滤波器,对非线性模型估计精度较差,分析比较两种常用的非线性滤波器,即UKF以及粒子滤波。UKF是基于高斯分布采样的估计模型,相比卡尔曼滤波会有很好的估计结果。粒子滤波是基于蒙特卡罗采样的推理模型,它不需要状态转移的先验信息,更利于羽毛球状态估计。(4)构建羽毛球机器人系统,包括视觉部分和机器人本体部分。视觉部分对机器人本体做位置全局校正和时间同步,羽毛球本体是击球动作执行单元。羽毛球机器人系统是一个高速运动且实时性较高的系统,建立了高动态下基于时间戳的位置校正模型,解决由于通信延时导致的位置同步精度下降问题。建立羽毛球挥拍模型和击球策略。
[Abstract]:Badminton is a sports event with high real-time requirement. In order to realize the human-machine match of badminton robot, it is necessary to track the badminton with high speed in real-time and dynamically. The problem of tracking dynamic targets exists widely in industrial production, life and military fields, such as missile interception system and road traffic real-time monitoring system. In this paper, the binocular vision system is used to capture the badminton in the dynamic flight process, and according to the characteristics of the badminton flight, the aerodynamics model of badminton is established. The main contents of this paper are as follows: (1) three kinds of dynamic object capture algorithms in common use are analyzed and compared: background difference method, optical flow method and inter-frame difference method. The advantages and disadvantages of these three algorithms are analyzed and compared. According to the flight characteristics of badminton, a badminton dynamic capture algorithm based on three-frame difference method and optical flow algorithm is established, and its advantages and disadvantages are analyzed. (2) the aerodynamics model of badminton is established. The effects of aerodynamics and gravity on badminton flying in the air are analyzed. The main factors affecting the flight trajectory of badminton are analyzed by experiments. The steady-state flight model and spin model of badminton are established by high-speed camera. (3) in the process of badminton trajectory prediction, there are model errors and measurement errors. The maximum likelihood estimation of badminton state is calculated by Bayesian filtering. Kalman filter is a linear filter with poor estimation accuracy for nonlinear models. Two kinds of nonlinear filters, namely UKF and particle filter, are analyzed and compared. UKF is an estimation model based on Gao Si distribution sampling. Compared with Kalman filter, there is a good estimation result. Particle filter is a reasoning model based on Monte Carlo sampling, which does not need prior information of state transition and is more conducive to badminton state estimation. (4) Badminton robot system is constructed, including visual part and robot body part. The vision part makes global position correction and time synchronization of robot body. Badminton body is the execution unit of batting action. Badminton robot system is a high-speed and real-time system. A position correction model based on timestamp under high dynamic condition is established to solve the problem of decreasing position synchronization precision caused by communication delay. Set up badminton swing model and batting strategy.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：G847;TP391.41

【参考文献】