一种改进RANSAC校验的失效卫星位姿估计回环检测方法

Loop closure detection for failure satellite position and attitude estimation with improved RANSAC

为解决失效卫星位姿估计中视觉SLAM方法回环检测效果差的问题,结合空间特殊的应用环境,提出一种改进RANSAC校验的空间失效卫星位姿估计回环检测方法。首先利用深度学习提取图像的全局特征和局部特征;然后对全局特征进行快速检索,建立回环候选帧集合;最后,将图像网格化并提取优质特征点作为RANSAC算法的随机抽样点集,采用改进的RANSAC算法对回环候选帧集合进行图像匹配校验,挑选出匹配程度最高的作为回环检测结果。实验表明:在牛津数据集上,所提方法与无校验的方法对比,100%准确率下的最大召回率提升了30.6%,说明改进RANSAC校验在回环检测中的有效性;在失效卫星数据集上,所提方法无需预训练,对光照变化的空间场景具有高适应性,和基于词袋模型(BoW)算法相比,缩减时间24.1%。

To solve the problem of poor loopback detection effect of visual SLAM method in position and attitude estimation of failed satellites,considering special spatial application environment,a loopback detection method with improved RANSAC check is proposed.Firstly,the global features and local features of the images are extracted.Then,the global features are compared quickly,and the loop candidates is established.Finally,images are gridded to screen high-quality features as the random sampling point set for RANSAC algorithm.The improved RANSAC algorithm is used to select the highest matched result.Experiments show that:on the Oxford dataset,the maximum recall at 100%precision of the proposed method is 30.6%higher than that of the method without check,indicating the effectiveness of the improved RANSAC in loop closure detection.On the failure satellite dataset,the proposed algorithm is highly adaptable to the space scene with illumination changes and requires no pre-training.Compared with BoW algorithm,the proposed method reduces the time by 24.1%.