Method for assessing the efficiency of using semantic information IN Visual SLAM

Abstract

The paper presents a method for assessing the efficiency of introducing semantic information into the ORB-SLAM3 algorithm. Non-determinism of the basic algorithm complicates an objective comparison of various modifications. In addition, existing semantic modifications are assessed on real data using image segmentation models, which introduces additional noise into the results. In order to overcome these limitations, a method is proposed that includes: a deterministic version of the algorithm that takes semantic information and information about dynamic objects as input; a special data set from robot motion sequences in an urban environment under various conditions, containing stereo pair frames, semantic masks and a list of moving objects; an algorithm for assessing changes in localization accuracy after introducing modifications. Based on the proposed method, an experimental assessment of various strategies for using semantic information in localization was carried out. The analysis revealed a significant influence of the quality of semantic segmentation on the accuracy of the algorithm. Methods for increasing the resilience to segmentation errors are proposed, including dynamic adjustment of semantic classes of map points and the use of information about object instances. In addition, it is shown that removing dynamic objects can either improve or worsen localization accuracy depending on the complexity of the environment.