基于ReLU神经网络的移动目标视觉伺服研究

针对移动目标跟踪以及抓取的问题,提出一种基于ReLU网络模型的单目视觉伺服系统。首先建立机器人视觉系统,完成对目标跟踪以及特征提取的任务,并通过结合单目视觉模型对其位姿进行估计,从而得到目标状态；然后,利用ReLU神经网络对机械臂的逆运动学进行学习,并用训练后网络模型构建单目视觉伺服系统的控制策略来避免机器人逆运动学求解计算量大、多解等问题；最后,为了提高抓取成功率,对末端执行器的运动轨迹进行规划。实验在NAO机器人平台上进行,根据实验结果证明方法的有效性。

Research of Visual Servo Based on ReLU Network

In regarding to the problem of moving target tracking and crawling, a monocular visual servo system based on ReLU network model is proposed. Firstly, the robot vision system is established for target tracking and feature extraction, and the target state is estimated from the monocular visual model; Then, the ReLU neural network is trained to construct the control strategy of the monocular visual servo system without large computational complexity and multiple solutions to the robot inverse kinematics; Finally, in order to increase the successful rate of crawling, the trajectory of the end effector is planned. The experiment is carried out on the NAO robot platform, and the results show the effectiveness of the method.