Research On 3D Vision Positioning Technology Of Indoor Mobile Robot

Mobile robot positioning technology is a key condition for realizing navigation tasks.Its purpose is to determine the exact position of the robot in the global map to achieve accurate pose tracking.The global self-localization process requires a pose search in a complete scene space.The amount of calculation is huge and the potential scene ambiguity and local scene dynamic changes in the localization process will lead to the failure of the a priori data association,resulting in a decrease in positioning accuracy Even global positioning fails.Two-dimensional image matching positioning technology can quickly determine the global pose of the robot,but the pose estimation accuracy is not high,and it is usually used in the initial stage of visual positioning.The three-dimensional point cloud registration positioning technology makes full use of environmental structure information and can obtain high-precision The pose of the robot is usually used in the precise positioning stage.Considering comprehensively,this paper proposes an indoor mobile robot three-dimensional visual navigation and positioning system combining two positioning technologies.It focuses on the improved algorithms in the initial and precise positioning stages and the solutions to the positioning problems,aiming to improve the accuracy and efficiency of positioning.The main work of this article is as follows:Firstly,during the initial positioning of robot vision,the traditional search dictionary-based image retrieval algorithm has a low search accuracy due to the low accuracy of image similarity calculation.This paper improves the computational robustness of image similarity through an improved visual dictionary constructed by a pyramid similarity matching strategy incorporating TF-IDF ideas.Aiming at the problem of global positioning failure,especially image ambiguity,setting up an autonomous motion search strategy to make the robot Explore the environment autonomously to obtain more available image matching information,thereby eliminating the problem of local convergence of poses.Secondly,the traditional point cloud registration algorithm has high time complexity and insufficient ability to deal with point cloud problems.This paper proposes a volume based on LORAX ideas and combining three effective improvement strategies of FC?CNN hierarchical acquisition,KDTree superpoint search and KDTree-ICP product neural network point cloud registration algorithm.This algorithm can extract more advanced point cloud features to improve the accuracy and speed of point cloud registration.The experiments show that the performance of the algorithm is better than the classic point cloud registration algorithm and can meet the needs of mobile robot real-time 3D positioning.Finally,this paper proposes a Monte Carlo precision positioning system based on improved visual dictionary image matching technology and convolutional neural network point cloud registration technology.In terms of particle weight calculation,the image matching model has high robustness but low efficiency.The point cloud registration model can effectively improve the calculation efficiency,stable performance and moderate robustness.In this paper,a point cloud registration model and image are designed.The adaptive observation model of the matching model is used as the weight calculation scheme;in the case of positioning failure recovery,the particle weight distribution change is analyzed by the enhanced Monte Carlo positioning(AMCL)algorithm and the particles are randomly sampled adaptively,using a combination of weight distribution change and continuous Frame matching is a response mechanism of two judgment strategies to solve the problem of robot abduction.