电子测量与仪器学报

孙龙龙,江明,焦传佳.基于运动矢量的改进视觉 SLAM 算法[J].电子测量与仪器学报,2020,34(9):23-31

基于运动矢量的改进视觉 SLAM 算法

Improved visual SLAM algorithm based on the motion vector

DOI：

英文关键词:simultaneous localization and mapping motion vector motion point detection loop closure detection

基金项目:国家自然科学基金(61271377)资助项目

作者	单位
孙龙龙	1. 安徽工程大学电气工程学院，2. 安徽工程大学高端装备先进感知与智能控制教育部重点实验室
江明	1. 安徽工程大学电气工程学院，2. 安徽工程大学高端装备先进感知与智能控制教育部重点实验室
焦传佳	1. 安徽工程大学电气工程学院，2. 安徽工程大学高端装备先进感知与智能控制教育部重点实验室

Author	Institution
Sun Longlong	1. School of Electrical Engineering, Anhui Polytechnic University,2. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education, Anhui Polytechnic University
Jiang Ming	1. School of Electrical Engineering, Anhui Polytechnic University,2. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education, Anhui Polytechnic University
Jiao Chuanjia	1. School of Electrical Engineering, Anhui Polytechnic University,2. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education, Anhui Polytechnic University

摘要点击次数: 565

全文下载次数: 902

中文摘要:

针对移动机器人运行场景中出现运动物体时,视觉同时定位与地图构建( SLAM)算法位姿估计误差大且构建地图不一致的问题,提出了一种基于特征点运动矢量的改进视觉 SLAM 算法。首先,引入基于特征点运动矢量的运动点检测算法。通过结合初始相机位姿,计算图像特征点的运动矢量,并使用期望最大化方法求解运动矢量角度的高斯混合模型参数,通过结合前一帧的运动点检测结果,从而区分当前图像中的运动特征点;其次,基于运动点检测结果,对当前帧相机位姿进行优化;再次,通过设置图像预处理环节,剔除运动点占比较大和与前一帧相似性较高的图像,提高闭环检测算法的计算效率;最后,使用剔除动态点后的图像特征点对场景进行描述,并改进单个节点处图像间相似性得分计算函数,经过闭环确认后,得到正确闭环。数据集实验表明,所提算法具有较高的位姿估计精度和较好的鲁棒性,同时能有效检测场景中闭环的存在,且建图效果较好。

英文摘要:

Aiming at the problem that the simultaneous localization and mapping ( SLAM) algorithm has a large pose error and inconsistent map construction when a moving object appears in mobile robot’ s operating scene, an improved visual SLAM algorithm based on feature point motion vector is proposed. Firstly, the algorithm of motion points based on feature point motion vector is introduced. The motion vector can be calculated by combining the initial camera pose, and the Gaussian mixture model parameters of its angle are solved by using the expectation maximization method. And the motion point detection result of the previous frame is used to distinguish motion points in the current image. Secondly, the camera pose will be optimized based on results of the motion point detection. Then the image is pre-processed, and images with a number of motion points and higher similarity to the previous frame will be eliminated, which can improve the calculation efficiency of loop closure detection. Finally, the scene is described by using feature points after excluding dynamic features, and the similarity score calculation function of two images at a single node is improved. After loop closure confirmation, the correct loop is obtained. The datasets experimental results show that the improved algorithm has better robustness and higher accuracy in the pose estimation. And it can effectively detect the existence of loops in the scene and has a good mapping.

查看全文查看/发表评论下载PDF阅读器