一种用于机器人抓取的单目实时三维重建系统

Real-Time Monocular 3D Reconstruction System for Robot’s Grasping

为支持家庭用服务机器人抓取任务,设计了一种针对家庭环境中常见物体的实时三维重建系统。系统假设相机在围绕物体运动的同时采集图像。首先借助从运动恢复结构算法求解相机在采集图像序列时的运动轨迹。根据运动轨迹中的位置信息,基于几何约束为相机运动轨迹中的每一帧选择最合适的配对图像。提出基于虚拟双目的立体匹配方法恢复每组配对中两帧图像共视区域的三维结构。利用L-R检查和抛物线拟合等后处理方法剔除误差较大的测点。最后用不同的图像配对计算出的三维结构生成最终的物体模型。实验结果表明生成的物体模型中位数相对误差在1.0%左右,满足机器人抓取任务中对物体模型的精度要求。同时系统平均运行速度超过25帧/秒,满足家庭用物体快速建模的要求。

A real-time3 D reconstruction system for common objects in home environment is proposed to provide support for grasping task of service robots.System assumes that the camera keeps moving and captures images in the same time.Trajectory of camera motion is first estimated by applying Structure from Motion algortithm.Based on position information of camera trajectory,most suitable matched image is selected for each camera frame under geometric constraints.Virtual stereo camera based motion stereo method is proposed to recover 3 D structure of covisible region for each two matched image pair.L-R check and parabola fitting and other post-processing method are applied to cull error measurements.3 D structure measurements from different image pairs are then used to generate object model.Experiments show median error of object 3 D model is about 1.0%and the accuracy meets the requirement of grasp task.Fast household object cosntruction requirement is also fullfilled as system process over 25 frames per second.