基于模型约束的人体姿态视觉识别算法研究

Research on human pose visual recognition algorithm based on model constraints

针对移乘服务机器人对人体姿态识别高精度性的要求,以及现有的人体姿态识别方法在关节遮挡情况下识别精度低的问题,提出了一种基于模型约束的人体姿态识别算法,以解决移乘操作前机器人系统对人体关节空间坐标的精确提取。首先采用OpenPose算法识别彩色图像中未遮挡关节的像素坐标,通过对RGB-D相机的彩色图像与深度图像进行对齐,将关节像素坐标转换为3D坐标。然后依据人体模型相关参数以及未被遮挡关节坐标计算与之相连接的被遮挡关节的空间坐标,用于提高遮挡关节的识别精度。实验结果表明,所提出的算法在关节未遮挡时识别精度为92%,在关节遮挡时达到了90%。单帧计算平均用时约为190 ms,满足移动服务机器人操作的实时性要求。

Aiming at the requirement of transfer service robots for high precision of human pose recognition and the problem of low recognition accuracy of existing human pose recognition methods under joint occlusion,this paper proposes a model constraint-based human pose recognition algorithm to solve the precise extraction of human body joint space coordinates of the robot system before the transfer operation.Firstly,the OpenPose algorithm is used to identify the pixel coordinates of the unoccluded joints in color image.Through aligning the color image and depth image of the RGB-D camera,the joint pixel coordinates are converted into 3 D coordinates.Then,the spatial coordinates of the occluded joint connected with the unoccluded joint are calculated according to the relevant parameters of the human model and the unoccluded joint coordinates,which are used to improve the recognition accuracy of the occluded joints.The experiment results show that the recognition accuracy of the proposed algorithm is 92%when the joint is unoccluded,and reaches 90%when the joint is occluded.The average time for a single frame calculation is about 190 ms,which meets the real-time requirements of transfer service robot operation.