Research On Indoor Robot Navigation Algorithm Based On VI-SLAM

The ability of indoor robots to accurately localise in unknown environments is a prerequisite for achieving autonomous navigation and completing various tasks.Although pure visual SLAM can achieve good localisation results,it is not effective in some scenes with high dynamics and low texture and high speed motion in a short period of time due to its limitations,and cannot meet the needs in specific scenarios.To this end,this paper investigates the localisation and navigation techniques for mobile robots in indoor complex environments based on visual inertial fusion algorithms.The research in this paper is as follows:1.The mathematical description of the visual SLAM problem and the theory of coordinate transformation and rotation in visual inertial fusion systems are analyzed,the imaging principle of the camera and the aberration model generated are described,and the characteristics of different types of cameras are introduced;the sources of IMU error are analyzed,and an IMU measurement model that takes random wandering error and Gaussian white noise into account is established,considering the large cumulative IMU error.2.Most of the traditional visual SLAM algorithms are studied based on the assumption of static scenes,which are prone to poor localization accuracy and stability in dynamic scenes.To address this problem,the ORB-SLAM3 algorithm is used as the basic framework,an independent target detection thread is added,the YOLOv3 target detection network is used for dynamic target classification,feature points located in the high dynamic target frame and not in the low dynamic target frame are eliminated,dynamic feature points and mis-matched feature points are further eliminated using the pair of polar geometric constraints and optical flow constraints,and finally the remaining feature The remaining feature points are then used for pose estimation.Through experimental validation on the TUM dataset,the absolute trajectory error of the system is reduced by more than 85% in high dynamic scenes compared to ORB-SLAM3,which greatly improves the localization accuracy.At the same time,the average time spent per image frame by the tracking thread is around 75 ms,which can meet the real-time requirements.3.For the pure vision system in high-speed motion or low-texture images are prone to tracking loss,poor positioning accuracy and other problems,designed a visual inertial tight coupling algorithm,first use IMU pre-integration strategy to complete the image frame to frame information alignment,IMU sensor’s high accuracy and high frame rate can compensate for the loss of visual information under high-speed motion,the use of visual residuals and IMU residuals to construct the objective optimization function,which is transformed into a least squares problem,the camera can solve the IMU drift problem in low-speed motion.For the estimated trajectory drift problem that occurs in large-scale complex environments,the similarity check of image frames and the repositioning of the positional estimate are performed on the closed-loop region of the trajectory using loopback detection based on the bag-of-words model Aiming at the problems of tracking loss and poor positioning accuracy in pure vision system in high-speed motion or low-texture images,a visual inertial close coupling algorithm is designed,first of all,the IMU pre-integration strategy is used to complete the information alignment between image frames,the high precision and high frame rate of the IMU sensor can make up for the lost visual information under high-speed motion,and the target optimization function is constructed by using the visual residual and IMU residual to convert it into a least squares problem,and the camera can solve the drift problem of the IMU in low-speed motion.For the estimation trajectory shift problem in large-scale complex environment,the loopback detection based on bag-ofword model is used to test the similarity of image frames and relocate the pose estimation on the closed-loop area of the trajectory.4.The Limo mobile robot model and scene model were established in the ROS environment,and the Gmapping algorithm was used to check the usability of the established model,but due to the computer configuration and other reasons,the established Limo model could not be used to check the visual inertial fusion algorithm.