用于行星探测车的视觉测程方法已经被广泛研究,但传统的视觉测程方法仅根据相邻的两帧影像进行运动估计,忽略了连续多帧影像间更强的几何约束。本文提出一种新的视觉测程方法,在连续多帧影像间进行特征追踪构成影像网,并用光束法平差进行运动估计。在特征追踪过程中,根据不同立体影像中欧氏距离的一致性剔除离群值,并自适应地选择几何关键帧。用蒙特卡罗方法分析了不同几何配置下视觉测程的精度。野外实验结果表明本文提出的新方法能有效提高探测车定位精度。

The use of Visual Odometry (VO) for planetary rover localization has been extensively researched. Traditionally, motion estimation is resolved only between adjacent frames; this neglected stronger geometric constraints possible from processing multiple consecutive frames. This paper presents a new visual odometry method based on bundle adjustment of an image network formed by feature tracking among multiple consecutive frames. In the feature-tracking process, outliers are detected based on the consistency of Euclidean distances in different stereo frames; geometric key frames are selected adaptively. Additionally, analysis of VO accuracy with respect to the geometric configuration of stereo camera is performed using Monte Carlo simulation. Results of field experiments demonstrate the effectiveness of the proposed method in improving localization accuracy.