Research On Visual SLAM Algorithm Based On Point Line Feature Fusion

With the continuous development of robot technology and intelligent perception technology,the importance of Simultaneous Localization and Mapping(Simultaneous Localization and Mapping,SLAM)in the mobile robot service industry is increasing.However,when faced with complex environments,visual SLAM based on feature points is prone to problems such as insufficient feature point information and uneven distribution,resulting in inaccurate system pose estimation.At the same time,in this kind of environment,there is often relatively rich line segment information,which can be used as line features to supplement the environmental feature information.Therefore,this paper improves the visual SLAM algorithm based on point-line feature fusion by adding the line features in the environment to the visual SLAM system.Experiments prove that the algorithm can effectively improve the positioning accuracy and robustness of the visual SLAM algorithm.Aiming at the performance degradation of the traditional point-based visual SLAM algorithm in weak texture environments,the algorithm first screens and optimizes blurred images to obtain image frames with better matching effects.Then use point feature information to judge whether to introduce line features,improve and optimize LSD(Line Segment Detector)line features through image pixel gradient density,and merge close line segments that meet the angle and coordinate thresholds to reduce the error caused by too many short-line features influence to improve subsequent matching quality.Finally,the progressive consistent sampling algorithm is adopted to eliminate the wrong matching points,improve the matching accuracy of the system feature points,and achieve the effect of improving the positioning accuracy.Aiming at the problem that mobile robots are prone to missing key frames in complex motion states,this paper adopts a key frame selection strategy based on motion states to supplement the key frames that may be missed when the robot moves in a large-angle curve to optimize the backend pose.Provide important original information and improve the utilization rate of scene information.In the closed-loop detection link,in order to ensure the uniformity of word distribution in the image,the Kmedians algorithm is used to cluster the features,and the weights are set reasonably according to the actual characteristics of different scenes.In order to distinguish the importance of different words,use the TF-IDF(Term Frequency-Inverse Document Frequency)model to calculate the weight of each word,and then judge the similarity between the image frames.When the similarity is high enough,it is judged that the two frames of images are closed loops,providing more pose constraints to further improve the composition accuracy of the system.The algorithm of this paper is verified by KITTI,TUM public data sets and real scene data sets such as offices,and the performance of different algorithms is compared.The experimental results show that the method in this paper can effectively improve the positioning accuracy of the system and the accuracy of closed-loop detection.