一种基于改进TF-IDF的SLAM回环检测算法

TF-IDF based loop closure detection algorithm for SLAM

提出了一种基于改进TF-IDF的视觉SLAM回环检测算法,用于检测曾经访问过的位置,来消除定位过程中的累积误差.首先,针对在人造建筑场景中使用SLAM算法对图像点特征进行计算容易导致检测失败的问题,采用图像中的直线作为特征来进行回环检测的计算.其次,在LBD(line band descriptor)图像线特征描述子的基础上进一步提取了二进制LBD描述子来进行视觉词典的构建,保证了线特征的处理效率.提出了一种改进的TF-IDF(term frequency&inverse document frequency)单词权重确定方法,提高了视觉单词评分之间的区分度.最后,以室内建筑环境和输电线路场景为例进行实验,结果显示,所提出的基于线特征的回环检测算法比基于点特征的算法有较高的检测准确率,有助于提高SLAM算法的计算性能.

A term frequency & inverse document frequency(TF-IDF) based loop closure detection algorithm for vision-based simultaneous localization and mapping(SLAM) was presented. The algorithm could detect the revisited places of the robot to eliminate the accumulated error during localization. Firstly, line features were used to detect the loop closure in the man-made environment, which is due to the point features could lead to a failure of SLAM. Secondly, the binary line band descriptor(LBD) was constructed based on LBD to build the vision dictionary, improving the efficiency of processing line features. Thirdly, a variant of TF-IDF was proposed to enhance the word discrimination. Finally, the experiments were carried out on indoor scene and transmission line scene.Experimental results show that the line-based algorithm outperforms point-based algorithm on the detection rate. It can help in ensuring the proformance of the SLAM computation.