基于kinect的改进RGB-D视觉里程计

PDF下载HTML

Title:

Improvement of kinect performance in RGB-D visual odometer

作者:

朱俊涛, 陈强

上海工程技术大学 电子电气工程学院，上海 201600

Author(s):

ZHU Juntao, CHEN Qiang

Electrical and Electronic Engineering College, Shanghai University of Engineering and Technology, Shanghai 201600, China

关键词:

kinect; 深度丢失; 融合算法; 特征点; ICP; PnP; 深度值; 位姿估计; BA优化模型; g2o

Keywords:

kinect; lack of depth; fusion algorithm; feature points; iterative closest point; perspective-n-point; depth value; pose estimation; BA optimization model; g2o

分类号:

TP242.6

DOI:

10.11992/tis.201903007

文献标志码:

A

摘要:

针对RGB-D视觉里程计中kinect相机所捕获的图像深度区域缺失的问题，提出了一种基于PnP(perspective-n-point)和ICP(iterative closest point)的融合优化算法。传统ICP算法迭代相机位姿时由于深度缺失，经常出现特征点丢失导致算法无法收敛或误差过大。本算法通过对特征点的深度值判定，建立BA优化模型，并利用g2o求解器进行特征点与相机位姿的优化。实验证明了该方法的有效性，提高了相机位姿估计的精度及算法的收敛成功率，从而提高了RGB-D视觉里程计的精确性和鲁棒性。

Abstract:

Kinect is a 3D camera that gives you the depth values associated with every pixel. It uses structured infrared light to determine depth values. Apart from these, you also have access to raw RGB-D data, and even the raw infrared data. Aiming to solve the problem of insufficient depth values for the images captured by Kinect camera in RGB-D visual odometer, we propose a fusion optimization algorithm based on Perspective-n-Point and iterative closest point (ICP). Because of the lack of depth values, traditional ICP algorithm often loses feature points when iterating the camera pose; this results in excessive error, or we can say that the algorithm is unable to converge. This algorithm establishes bat algorithm optimization model by judging the depth of feature points and optimizes the feature point of poses and camera using g2o solver. Experiments show that the method is effective and improves the accuracy of camera pose estimation and the convergence success rate of the algorithm, thus improving the accuracy and robustness of RGB-D visual odometer.