基于YOLOv3和EPnP算法的多药盒姿态估计

Multi-pillbox Attitude Estimation Based on YOLOv3 and EPnP Algorithm

摘要:

针对机械臂药盒抓取操作中对药盒定位和姿态估计的要求,提出一种基于YOLOv3深度学习算法和EPnP算法相结合的多药盒姿态估计方法,此方法主要分为多药盒定位和姿态估计两部分。首先通过YOLOv3算法实现药盒的快速精确定位,并通过定位框分割出单个药盒。然后进行特征提取和特征匹配并估计单应矩阵。通过对单应矩阵的透视矩阵变换求得药盒平面四个角点的像素坐标并作为EPnP求解所需的2D点,结合药盒先验尺寸信息在相机坐标系下构建药盒对应的3D点坐标以实现药盒姿态求解。通过结合OptiTrack系统设计了药盒姿态精度对比实验,结果表明,本算法充分发挥了YOLOv3算法兼具快速性和准确性的优势,并且具有良好的姿态估计精度,总体算法速度达到15FPS,药盒姿态估计平均误差小于0.5度。

Abstract:

A multi-pill box attitude estimation method based on the YOLOv3 deep learning algorithm and the EPnP algorithm is proposed to deal with requirements for location and attitude estimation of pill boxes in the manipulator-gripping operation for pill boxes. This method is constructed by the multi-pill box location part and the attitude estimated part. First, the YOLOv3 algorithm is used to achieve the fast and accurate location of pill boxes, and located pill boxes are distinguished by location boxes. Then, feature extraction and feature matching are performed and the homography matrix is estimated. Pixel coordinates of four corner points in the pill-box plane are obtained by transforming the perspective matrix of the homography matrix, and these pixel coordinates are used as 2D points for EPnP solution. 3D coordinates of pill boxes are constructed under the camera frame by using pill boxes’ prior size information to compute postures of pill boxes. A comparative experiment for box attitude accuracy is designed based on the OptiTrack system, and experimental results show that the algorithm fully utilizes the advantages of the YOLOv3 algorithm e.g., high speed and accuracy, and has good attitude estimation accuracy. The overall computational speed reaches 15FPS, the estimated average error is less than 0.5 degree.