Research On Multi-view Complementarity Based SLAM

Simultaneous Localization And Mapping(SLAM)has become a hot topic in the field of Artificial Intelligence.SLAM is the process that intelligent agents compute their own poses and the 3D maps of surrounding environments,based on the data acquired through the sensors mounted on the agents.SLAM is widely used in automatic driving,autonomous mobile robots,virtual reality,augmented reality and other fields.For example,autonomous driving cars can locate and construct the surrounding scenes based on SLAM,to avoid obstacles and plan routes.Domestic service robots need to localize themselves and reconstruct scene maps,in order to plan the path and interact with the environments.As a branch of SLAM,Visual SLAM takes sequential images of scenes as input and outputs the camera pose at each frame and the 3D map of the sensed environment.At present,visual SLAM technologies have been widely studied.However,there are still some shortcomings as follows:1.Existing SLAM algorithms directly utilized features extracted from sequential images in bundle adjustment.However,the detections often contain noises,which will interfere with the optimization and deduce the accuracy in localization.2.Most SLAM algorithms paid more attention to the localization accuracy,but their performance in mapping is limited.For example,the line-based SLAM algorithms generally reconstruct 3D maps with incomplete and unstructured lines.3.Many scholars integrated different features,e.g.line and points,into SLAM systems,to deal with challenging scenes.However,these systems still treat each type of features separately,while neglecting the relation between them.After comprehensive analyses on the above issues,we develop a SLAM framework supporting multi-view complementarity.Compared with the traditional SLAM framework,this framework makes full use of the temporal information and geometric relationship between 2D observations.Based on this framework,we carry out works as follows:1.Aiming at solving the problem that existing SLAM algorithms cannot effectively suppress unreliable features,we propose to use multi-view information to measure the credibility of 3D features and suppress the negative impact from unreliable data.Experimental results show that the credibility-based SLAM obtains higher accuracy than the state-of-the-arts.2.Aiming at solving the problem of low quality of 3D line maps,we propose to construct the matched line features into a sequence,and optimize the end points of the sequence after pose optimization.This helps improve the accuracy of the end points,and thereby improving the integrity of the 3D maps.Through the reconstruction results of multiple scenes,we demonstrate the proposed SLAM algorithm with sequential line optimization is effective in improving mapping quality.3.Aiming at solving the problem that the relation between hybrid features was ignored,upon our framework of multi-view complementarity,we establish the structural connection between hybrid features.In bundle adjustment,the hybrid features with structural connection are used to optimize the 3D maps.Experiments show that the quality of reconstructed maps has been further improved.