基于相似度函数的上肢关节点改进阈值滤波方法及其应用

Method and application of the improved upper limb joint point-threshold filtering based on similarity function

针对视觉传感器受自身硬件条件及环境扰动的影响,采集到的关节点运动数据往往存在异常值的问题,为了提高Kinect获取的上肢关节点运动数据的准确性,提出了一种基于相似度函数的上肢关节点改进阈值滤波方法,来减小异常值对运动追踪的影响。基于前景轮廓相似度、肢体比例相似度和分布相似度构建了关节点相似度函数,将其用于评估截取的每帧关节点数据的可信度;分析了函数中各项参数对整体相似度的影响,基于相似度函数改进了阈值筛选法,并将其用于评估后的关节点数据的滤波;基于滤波后的上肢关节点运动数据完成了人机交互实验,提高了机械臂模仿人臂抓取的成功率。在跟踪人体上肢关节点运动时,与传统滤波方法相比,所提方法对关节点滤波的误差率更低,且具有较好的滤波稳定性和滤波速度。

As the vision sensor is affected by its own hardware conditions and environmental disturbances, the collected joint point motion data often has the problem of abnormal values. On the basis of this phenomenon, this paper designs an improved threshold filtering method based on the similarity function of the upper limb joint point. This new method can reduce the influence of abnormal values on motion tracking and improve the filtering accuracy of the upper limb joint point motion data obtained by Kinect. The joint point similarity function is constructed based on the similarity of foreground contour, the similarity of limb proportion and the similarity of distribution, and it is used to evaluate the reliability of the intercepted joint point data of each frame. Based on the similarity function, the threshold screening method is improved, and it is used to filter the joint point data after evaluation. Based on the filtered upper limb joint point motion data, the human-robotics interaction experiment is completed and shows a good grasping success rate of robot arms. Compared with traditional filtering methods, this method has a lower error rate for filtering joint points, and has better filtering stability and filtering speed when tracking different movements of the upper limbs.