Research On Visual Inertia SLAM Based On Point And Line Feature Fusion

With the increasingly widespread application of Simultaneous Localization And Mapping(SLAM)technology in the fields of autonomous driving and robotics,a single sensor can no longer meet the positioning and mapping requirements in actual environments.SLAM technology based on multi-sensor information fusion has become the main development direction.The visual inertia system composed of camera and IMU data fusion has higher positioning accuracy and robustness,but the visual inertia SLAM system based solely on point features performs poorly in complex environments such as low texture and drastic changes in lighting,making positioning easy to lose.Line segment features are widely present in structured environments and can be used as supplementary features for robot localization and mapping.However,the existing visual inertia SLAM schemes for point line feature fusion have problems in practical environments such as high computational complexity and poor real-time performance in feature extraction,and the positioning accuracy also needs to be improved.In response to the above issues,this article will improve and validate the algorithm from the following aspects:Firstly,to improve the adaptability of SLAM systems to the environment and reduce redundant features in actual environments,a dynamic extraction strategy for point and line features is proposed.Calculate the number of image frame feature points,inter frame speed,and other information,and introduce them into the front-end line feature processing thread to classify the actual scene,dynamically select point line feature extraction strategies,and thereby improve the SLAM system’s adaptability to the environment.At the same time,the frequency of line feature extraction is controlled in real-time based on multi-source information,reducing system computational complexity,improving algorithm real-time performance,and achieving high positioning accuracy with a small number of line features.Then,in response to the problem that existing wire feature extraction algorithms are prone to generating a large number of useless short line features,and to improve the accuracy of line segment feature matching,a line segment feature optimization strategy based on conditional filtering is proposed.Filter the length of line segments in the image to remove inferior short line features.In addition,in the actual environment,the matching and tracking of vertical line features have higher mutual difference.Therefore,the algorithm in this paper will filter the angle of line segment features,give priority to vertical lines for feature matching,and improve the positioning accuracy and robustness of SLAM system.Secondly,to further improve positioning accuracy,this article optimizes the backend sliding window algorithm.By setting the window threshold to make its size more suitable for the SLAM proposed in this article,increasing the amount of visual and inertial data within the sliding window,and completing key frame filtering and edge operation,the degree of data association is strengthened,and the system positioning accuracy is improved.Finally,simulation analysis and real scene verification.By using a publicly available dataset and comparing it with the traditional point line feature fusion visual inertia SLAM algorithm,the improvement effect of the SLAM system proposed in this paper in terms of positioning accuracy and real-time performance is verified.In addition,establish a visual inertial system testing platform to calibrate the two sensors separately and jointly.And design indoor and outdoor actual testing scenarios to verify the positioning accuracy and mapping effect of the SLAM system proposed in this article.