体感操控多臂棚室机器人作业决策规划算法研究

发布时间：2018-03-29 16:43

本文选题：多臂温室机器人　切入点：决策规划　出处：《农业机械学报》2017年03期

【摘要】：针对目前棚室内机器人作业分析算法智能性不足、准确作业率较低,且一次巡航过程只能进行单一作业,存在使用效率不高的问题,提出了一种搭载在三臂棚室机器人上,基于体感操控作业的决策规划算法,用Kinect采集含操作人员位姿信息的深度图像,结合随机森林统计学习理论和基于高斯核函数的Mean shift算法,确定了代表人体位姿的20个关键骨骼点坐标,在此基础上提出了一种基于模式切换的三臂映射关系,将骨骼点信息映射到机器人工作空间,使人的两只手臂能自如的控制三臂机器人,在一次巡航中完成多种棚室作业;此外,还提出了一种结合骨骼追踪技术和YCb Cr颜色空间的手势特征分割方法,实现了用手势控制机器人末端执行器作业。最后,搭建了用于测试体感决策算法的三臂机器人样机,进行了针对该决策算法的精确性试验,根据试验误差数据对肩部关节夹角采用离散化取值识别,解决了肩部关节识别误差,结果表明:测试者被捕捉到的关节处夹角和机器人对应关节夹角的最大映射误差为1.90°,上位机发送夹角值与机器人实际转动的夹角值最大误差为0.80°,在误差允许范围内,同时在该精度下完成一套采摘加喷施作业指令,平均耗时13.34 s,且操作者还可通过体感操控训练进一步提高机器人作业性能,表明该算法具有准确性和实用性。
[Abstract]:Aiming at the problem that the analysis algorithm of indoor robot in greenhouse is not intelligent enough, the rate of accurate operation is low, and only a single operation can be carried out in one cruise process, and there is not high efficiency in use, a kind of robot mounted on a three-arm shed room robot is put forward in this paper. Based on the decision planning algorithm of somatosensory manipulation, the depth images of operators' position and pose information are collected by Kinect, combined with stochastic forest statistics learning theory and Mean shift algorithm based on Gao Si kernel function. The coordinates of 20 key skeletal points representing human posture are determined. Based on this, a three-arm mapping relationship based on pattern switching is proposed, which maps the information of skeletal points to the robot workspace. In addition, a method of hand gesture feature segmentation combining bone tracking technology and YCb Cr color space is proposed. Finally, a prototype of a three-arm robot used to test the body sense decision algorithm is built, and the precision test of the decision algorithm is carried out. According to the experimental error data, the angle of shoulder joint is recognized by discretization, and the recognition error of shoulder joint is solved. The results show that the maximum mapping error between the joint angle captured by the tester and the corresponding joint angle of the robot is 1.90 掳, and the maximum error between the transmitting angle of the upper computer and the actual rotation angle of the robot is 0.80 掳. At the same time, a set of picking and spraying instructions is completed with this precision, which takes an average time of 13.34 seconds, and the operator can further improve the performance of the robot through the training of body sense control, which shows that the algorithm is accurate and practical.
【作者单位】：东北农业大学工程学院;
【基金】：黑龙江省普通高等学校青年创新人才培养计划项目(LR-356214) 国家自然科学基金项目(51405078) 黑龙江省博士后基金项目(LBH-Z13022) 东北农业大学“青年才俊”项目(518020)
【分类号】：S625.3;TP212;TP242

【参考文献】