基于改进的YOLOv3与EPnP算法的几何体位姿估计

Pose estimation of geometry model based on improved YOLOv3 and EPnP algorithms

针对机器人动态几何体模型抓取操作中存在对模型检测和位姿估计的问题,提出了一种基于改进的YOLOv3算法和EPnP算法相结合的几何体模型位姿估计方法。首先为提升YOLOv3算法的检测效果,采用完全交并比损失函数(CIOU)计算位置回归损失,同时引入通道注意力机制(ECAnet),运用网络自适应性注意重要信息,调整特征图。其次预测几何体模型3D虚拟控制点的2D像素坐标,利用EPnP算法求解目标物体的位姿。实验结果表明,与传统的YOLOv3算法相比,改进的YOLOv3算法mAP和F1性能分别提高了约6%和3%,并且能够较好的得到模型的中心和3D边框尺寸,有较好的位姿估计精度,验证了该方法的可行性和有效性。

Aiming at the problems of model detection and pose estimation in the grasping operation of robot dynamic geometry model,a pose estimation method of geometry model was proposed based on the combination of improved YOLOv3 algorithm and EPnP algorithm.Firstly,in order to improve the detection effect of YOLOv3 algorithm,CIOU function is used as position regression loss function,and on this basis,ECANet attention mechanism is introduced,and network adaptability is used to pay attention to important information and adjust the feature graph.Secondly,2D pixel coordinates of 3D virtual control points of the geometry model are predicted,and EPnP algorithm is used to solve the pose of the target object.Experimental results show that compared with the traditional YOLOv3 algorithm,the mAP and F1 performance of the improved YOLOv3 algorithm is improved by 6%and 3%respectively,and the center and 3D border size of the model can be obtained better,and the pose estimation accuracy is better,which verifies the feasibility and effectiveness of the proposed method.