最小化光度误差先验的视觉SLAM算法

Visual SLAM Algorithm for Minimizing Photometric Error Prior

在同时定位与地图构建(Simultaneous Localization and Mapping,SLAM)中,传统基于点特征的方法存在精度不足、剧烈抖动容易跟踪失败等问题,从而降低系统性能.针对此问题,本文提出一种新的视觉SLAM算法.首先使用双线性插值法得到特征点的灰度值,根据两帧之间的最小化光度误差得到当前帧估计位姿.其次为进一步降低位姿估计的误差,通过最小化重投影误差优化当前帧位姿,提高相机定位精度.最后为改善建图精度,提出一种新的关键帧选择机制,通过前端得到的位姿来衡量上一关键帧与当前帧的运动幅度,根据运动幅度判断当前帧是否加入关键帧序列,合理化关键帧选择方法.本文采用TUM数据集进行实验,与ORB-SLAM2相比,本文方法降低了相机的定位误差,提高了系统的鲁棒性.

In Simultaneous Localization and Mapping(SLAM),there are many challenges in traditional point feature-based approaches,such as insufficient accuracy,motion jitter and tracking failure,w hich reduce the performance of the algorithms.This paper proposes a novel algorithm called visual SLAM to handle these problems.First,w e present the bilinear interpolation method to get the gray values of the feature points,the estimated pose of the current frame can be obtained by minimizing the photometric error betw een tw o frames.Second,in order to further reduce the pose error,the pose of the current frame is optimized by minimizing the reprojection error,so as to improve the localization accuracy.Finally,w e come up w ith a new keyframe selection mechanism to improve the mapping accuracy,w hich measures the motion amplitude from the previous key frame to the current frame according to the pose,and judges w hether the current frame is added to the keyframe sequence to rationalize the selection of keyframes.This paper thoroughly evaluates the system on the TUM RGB-D benchmark and compares it w ith ORB-SLAM2,our method has better accuracy and robustness.