一种基于点云实例分割的六维位姿估计方法
网络安全与数据治理
周剑
苏州深浅优视智能科技有限公司
摘要: 提出了一种基于SoftGroup实例分割模型和PCA主成分分析算法来估计物体位姿的方法。在工业自动化领域,通常会为诸如机器人、机械臂配备视觉系统并利用二维图像估算目标物体位置,但当目标物体出现堆叠、遮挡等复杂场景时,对二维图形的识别精度往往有所下降。为准确、高效地获取物体位置,充分利用三维点云数据的高分辨率、高精度的优势:首先将深度相机采集到的RGB-D图像转为点云图,接着利用SoftGroup模型分割出点云图中的目标对象,最后用PCA算法得到目标的六维位姿。在自制工件数据集上进行验证,结果表明对三种工件识别的平均AP高达97.5%,单张点云图识别用时仅0.73 ms,证明所提出的方法具有高效性和实时性,为诸如机器人定位、机械臂自主抓取场景带来了全新的视角和解决方案,具有显著的工程应用潜力。
中圖分類號:TP391文獻標識碼:ADOI:10-19358/j-issn-2097-1788-2024-05-006
引用格式:周劍.一種基于點云實例分割的六維位姿估計方法[J].網(wǎng)絡安全與數(shù)據(jù)治理,2024,43(5):42-45,60.
引用格式:周劍.一種基于點云實例分割的六維位姿估計方法[J].網(wǎng)絡安全與數(shù)據(jù)治理,2024,43(5):42-45,60.
6D pose estimation based on point cloud instance segmentation
Zhou Jian
DEEPerceptron Tech
Abstract: This paper proposes a method based on the SoftGroup instance segmentation model and Principal Component Analysis (PCA) algorithm for estimating object poses. In the field of industrial automation, visual systems are often equipped on robots and robotic arms to estimate the position of target objects using 2D images. However, in complex scenarios such as stacking and occlusion, the recognition accuracy of 2D images tends to decrease. To accurately and efficiently obtain object positions, this paper fully leverages the high-resolution and high-precision advantages of 3D point cloud data. Firstly, RGB-D images captured by a depth camera are converted into point cloud images. Then, the SoftGroup model is employed to segment the target objects in the point cloud image, and finally, the PCA algorithm is used to obtain the six-dimensional pose of the target. Validation on a self-made dataset shows an average AP of 97.5% for the recognition of three types of objects. The recognition time for a single point cloud image is only 0.73 ms, demonstrating the efficiency and real-time capability of the proposed method. This approach provides a new perspective and solution for scenarios such as robot localization and autonomous grasping of robotic arms, with significant potential for practical engineering applications.
Key words : point cloud data; SoftGroup instance segmentation; 6D pose estimation
引言
近年,隨著激光掃描儀、相機、三維掃描儀等硬件設備的發(fā)展與普及,點云數(shù)據(jù)的獲取途徑變得更加多樣,數(shù)據(jù)獲取的難度不斷降低。相較于二維圖像,三維點云數(shù)據(jù)具備無可比擬的優(yōu)勢。其高分辨率、高精度、高緯度的特性賦予點云數(shù)據(jù)更為豐富的空間幾何信息,能夠直觀地表達物體的形狀特征。近年來,點云數(shù)據(jù)在工業(yè)測量、機械臂抓取、目標檢測、機器人視覺等領域得到了廣泛應用[1–3]。
在工業(yè)自動化領域,通常需要先獲得物體的位姿信息再進行后續(xù)抓取動作。自動抓取物體可分為結(jié)構(gòu)化場景和非結(jié)構(gòu)化場景。在結(jié)構(gòu)化工作場景中,機械臂抓取固定位置的物體,該模式需要進行大量調(diào)試和示教工作,機械臂只能按照預設程序進行工作,缺乏自主識別和決策能力,一旦目標物體發(fā)生形變或位置偏移,可能導致抓取失?。辉诜墙Y(jié)構(gòu)化場景中,通常為機械臂配備視覺感知硬件和目標檢測算法,以使機械臂能夠感知并理解相對復雜的抓取環(huán)境。然而,在實際復雜的抓取場景下(如散亂、堆疊、遮擋),常見的目標檢測方法如點云配準[4]、二維圖像實例分割[5]的精度有所下降,從而影響抓取效率[6]。
本文詳細內(nèi)容請下載:
http://www.ihrv.cn/resource/share/2000006014
作者信息:
周劍
(蘇州深淺優(yōu)視智能科技有限公司,江蘇蘇州215124)

此內(nèi)容為AET網(wǎng)站原創(chuàng),未經(jīng)授權(quán)禁止轉(zhuǎn)載。
