Research On Simultaneous Localization And Map Construction Of Vision-based Mobile Robots

Simultaneous Localization and Mapping(SLAM)is the technical foundation for mobile robots to achieve more autonomous and intelligent behavior.Vision sensors can meet the low-cost positioning and navigation needs of indoor mobile robots.This paper is devoted to the SLAM algorithm of mobile robots based on the vision sensor.The main research work is as follows:(1)The localization algorithm based on the feature point method is studied.Aiming at the disadvantage that the traditional ORB algorithm is easily affected by changes in brightness and texture,an adaptive threshold feature extraction method is proposed.Firstly associate the threshold with the pixel brightness information,and then dynamically adjust the threshold by adding an activation function controlled by the pixel contrast to the threshold to adapt to the change in the texture of the scene.This method can also effectively extract feature point information in light changes and weakly textured scenes.(2)Based on the principle of regional motion consistency,an RMC-based mismatch removal method is proposed,which uses motion characteristics and pixel characteristics to achieve coarse and fine removal of mismatched point pairs,respectively.This method is superior to RANSAC in both speed and stability.The single frame matching time is increased by more than 40%,and the number of pairs of correct matching points retained is increased by more than 30%.(3)A new key frame selection strategy is proposed to solve the tracking loss problem that robots may have.The temporal distance,spatial distance,and inter-internal point information are used as the judgment conditions for key frame selection.In addition,an optimal candidate key frame judgment method is designed to reduce the probability of tracking loss and improve the robustness of the system.(4)Aiming at the problem of low detection rate of the loop closure detection method based on the bag-of-words model,a loop closure detection method fusing semantic information is proposed.This method makes use of abstract semantic information to make up for the shortcomings of the bag-of-words model.It can not only improve the similarity contrast between loopback and non-loopback scenes,make loopback easier to judge,but also expand the scope of loopback scene detection and improve the recall rate.Provide constraints for global optimization.(5)In the indoor environment,the construction of the 3D point cloud map and the 3D octree map of the real laboratory scene is completed,which verifies the effectiveness of the map construction algorithm in this paper.