Research On SLAM Algorithm For Mobile Robot Based On Vision And IMU Fusion

Simultaneous Location And Mapping(SLAM)refers to a mobile robot that collects physical information from an experimental scene during its motion in an unknown environment by using multiple sensors carried by itself to build a model of the environment and estimate its own motion information at the same time.SLAM algorithms are divided into two main categories,laser SLAM and vision SLAM algorithms,depending on the type of sensor used by the mobile robot.Laser generators have not been widely used in the field of SLAM due to the limitations of many factors of their own,such as high cost and high power consumption,while visual SLAM,which is more cost-effective and has greater scene utilization,has gradually become the mainstream of autonomous localization and mapping research.However,the existing vision SLAM algorithm extracts more redundant feature information when processing images in dense texture scenes,which affects the system’s real-time performance,makes it difficult to provide a better source of location information for closed-loop detection and optimization in complex environments due to improper selection of keyframes,and causes poor image quality due to problems such as excessive viewpoint difference or long shutter time when using only frame image information obtained from a single vision sensor,which results in inaccurate pose estimation and makes the constructed trajectory map inconsistent with the real trajectory.In this article,we proposed an improved keyframe extraction strategybased Fast PL-SLAM algorithm(IFPL-SLAM)to address the problems that the existing PL-SLAM(Point and Line Simultaneous Localization and Mapping)algorithm based on point and line features for pose estimation extracts and matches point and line features simultaneously for pose solution in a dense texture information scene,which has redundancy in feature calculation and affects the real-time performance of the system,as well as the problems that the existing method is prone to inaccurate pose estimation and incomplete closed-loop correction due to the omission of keyframe selection during the curved motion of the mobile robot.The algorithm adopts a combined point and line feature-based positional estimation strategy at the front end of the system to construct a local map of the environment,evaluates the point features extracted from the image by information entropy,and decides whether to call the line features in the environment for fused pose estimation based on the calculated information amount of point features,which effectively avoids the waste of computational resources caused by the simultaneous use of point and line features in dense texture scenes.At the same time,to address the problem that mobile robots tend to miss keyframes near curve motion inflection points,this article adopts a keyframe index selection strategy combining cis and trans directions to select keyframes and supplement keyframes that may be missed by the cis index near curve motion inflection points,which provides important original information for SLAM back-end pose optimization and improves the utilization of scene information near curves.In the closed-loop detection link,the closed-loop detection algorithm based on the visual bag of words detects whether the current scene is a familiar scene,and when the mobile robot arrives at the familiar scene,the accumulated error caused by the long operation is eliminated by the a priori bit pose of the familiar scene.And on this basis,this article addresses the problem that the image frame information acquired using a single vision sensor may have poor image quality due to problems such as excessive viewpoint difference or long shutter time,and the inaccurate pose estimation makes the constructed trajectory map inconsistent with the real trajectory.The accuracy of the mobile robot’s pose estimation is improved by the key means of adding complementary IMU inertial sensors.The performance of the algorithm proposed in this article is verified and analyzed by publicly available KITTI,TUM,Eu Ro C and real scene datasets.The results demonstrated that the method in this article effectively improves the system real-time performance,localization accuracy and the accuracy of closed-loop detection.