Data Association Research For Visual SLAM

Simultaneous localization and mapping(SLAM)is used to solve the positioning and mapping problems of robots in unknown environments.It is a key technology to realize autonomous positioning of robots,and it is of great significance for robots to step into practical applications.In recent years,relying on the advantages of low cost of cameras and abundant information,visual SLAM technology has become a research hotspot.Data association is the core technology of SLAM system,which aims to find the correspondence between the observation and the environment.Stable and accurate data association is the premise of SLAM's accurate positioning and mapping.The purpose of this paper is to improve the accuracy of data association,from the perspective of line features and saliency maps.The existing feature based SLAM mainly relies on uniformly extracting points from the scene,then establish association with the environment.But there are two main problems: On the one hand,sparse point features can easily lead to data association failures.But for the scene such like corridors and structured roads,although they lack point features,but there are rich line features.Therefore,research on data association algorithms based on point and line features is meaningful.Line features can be used to complete data association when point features are scarce;On the other hand,there are some high-saliency areas in the scene.The feature points in these areas are more robust to changes in perspective and lighting.A region selection mechanism can be established.Select feature points that make it easier to achieve correct data association.Based on this background,this article proposes the following innovations:1.The existing point and line SLAM algorithm completely separates the processing of point and line features.To break this tradition,the line feature association method based on the point and line invariants is introduced.Line matching is completed directly by using the feature point matching results,which strengthens the coupling of the point and line features,and reducing the system's resource consumption for line matching;2.Aiming at the phenomenon that the line feature matching accuracy is generally lower than the point feature matching accuracy,when constructing the system reprojection error,the points and line feature errors are weighted according to the density of the feature points.This data association method based mainly on point features and supplemented by line features,which can avoid the negative impact of low line matching accuracy;3.The existing way to extract point features is uniform extraction.In order to make this process more efficient,a saliency map-based feature extraction scheme is introduced.The saliency map is obtained through the saliency and semantic information of the input image,and used to design the extraction position of the point features to ensure that the salient areas extract as many point features as possible.Experiments show that both improved data association methods can improve the accuracy and robustness of system tracking,and ensure good real-time performance.