基于特征点改进的视觉SLAM定位研究

Research on Improved Visual SLAM Localization Based on Feature Points

摘要:

为改善视觉SLAM(SimultaneousLocalizationandMapping)系统在低纹理环境下定位精度较低的现象,提出一种改进的ORB(OrientedFASTandRotatedBRIEF)特征点提取策略和一种关键帧选择机制；首先采用多尺度分析和基于局部灰度的特征检测方法克服一般ORB算法缺乏尺度和旋转描述的缺点；其次提出一种基于高斯模糊的图像信息增强方法解决传统ORB特征点提取方法在纹理信息不被突出环境下容易失效的问题,并对图像进行象限分割使特征点均匀分布；最后为剔除劣质关键帧,设计了一种综合时间因素与特征点数量因素的关键帧选择机制；将提出的方法移植到ORB_SLAM2上,并在TUM数据集上测试,实验结果表明,视觉SLAM系统的定位误差平均降低14.688%,证实了本文方法的有效性。

Abstract:

In order to improve the low localization accuracy of visual simultaneous localization and mapping (SLAM) system in low texture environment, an improved oriented fast and rotated brief (ORB) feature point extraction strategy and a keyframe selection mechanism are proposed. Firstly, multi-scale analysis and feature detection method based on local gray level are used to overcome the shortcomings of general ORB algorithm which lacks scale and rotation description. Secondly, an image information enhancement method based on Gaussian blur is proposed to solve the problem that the traditional ORB feature point extraction method is easy to fail in the environment where the texture information is not prominent, and the image is segmented to make the feature points evenly distributed. Finally, in order to eliminate inferior keyframes, a keyframe selection mechanism combining time factor and feature point number factor is designed. The proposed method is transplanted to ORB _ SLAM2 and tested on the TUM dataset. The experimental results show that the localization error of the visual SLAM system is reduced by 14.688 % on average, which confirms the effectiveness of the proposed method.