基于双目视觉的物体识别定位与抓取

Object Recognition and Capture Based on Binocular Vision

为解决目标识别精度低、定位与抓取配合困难问题,通过对算法的改进研究了物体的识别定位与抓取,包括识别目标物体、对于目标物体形心的位置三维重建、对于目标物体的姿态估计、对于手眼标定后的抓取4个方面。在目标识别中,首先通过快速鲁棒特征(speeded up robust features,SURF)算法结合Grabcut算法提取出目标。在位置求解中,利用模板匹配求取目标物体形心的世界坐标。在姿态估计中,算法的流程为:利用匹配点对求取左图中物体母线斜率,再随机取斜率等于左图母线斜率的两点,通过两点的世界坐标求出目标物体的姿态。在抓取中,采用的眼在手上,先建立工件坐标系,再进行坐标转换,通过机器人参数求得逆解。研究结果表明:误差均在较小范围内且机器人可在有效工作范围实现抓取。可见算法可靠以及整体实验的正确性。

In order to solve problems of low accuracy of target recognition and difficult in coordination of location and grabbing, an improved algorithm was used to investigate object recognition, location and grabbing. It was divided into four aspects, target object indentification, three-dimensional reconstruction of the object’s centroid, posture estimation of the target object and target grasping after hand-eye calibration. In target recognition, the target was extracted by speeded up robust features(SURF) algorithm combined with Grabcut algorithm. In the position solution, the world coordinate of the object’s centroid was obtained by template matching. In the process of posture estimation, the matching point pair were used to obtain the slope of the object generatrix in the left picture firstly, randomly taking two points with the slope equal were taken to the slope of the left picture generatrix, and finally the posture of the target object were obtained through the world coordinates of the two points. In the process of grabbing, the way of eye on hand was adopted. The workpiece frame was established, and the coordinate transformation was carried out, then, the inverse solution was obtained through the robot parameters. The results show that the errors are in a small range and the robot can grasp in the effective working space. It is concluded that the algorithm is reliable and the whole experiment is verified to be correct.