| In recent years,robot technology and computer vision technology have been continuously developed.As a cross technology between the two,vision SLAM system has been widely used in all aspects of life,such as virtual reality,automatic navigation,three-dimensional reconstruction,and so on.How to improve the accuracy,real-time and robustness of visual SLAM systems has become a research hotspot for scholars.This paper mainly studies the closed-loop detection module and back-end optimization module of the visual SLAM system.The closed-loop detection of the visual SLAM system is mainly based on the bag-of-words model,and the visual dictionary is the basis of the bag-of-words model.Its representation ability affects the accuracy of the closed-loop detection algorithm,and its scale seriously affects the operation efficiency of the closed-loop detection algorithm.This paper proposes a fusion dictionary expansion algorithm.First,a basic visual dictionary is constructed offline,and then it is expanded online.While ensuring the real-time performance of the algorithm,it reduces the feature quantization error and improves the dictionary’s representation ability..Continuous consistency detection is a technical method used to improve the accuracy of closed-loop detection,but it will not be able to detect the closed-loop that may exist at the end of the image sequence.This paper proposes an improved closed-loop detection strategy,which uses consistency detection to ensure the accuracy of the closed-loop,and at the same time at the end of the image sequence,raises the threshold value of the feature point matching detection to perform another detection of the closed-loop candidate at the end to avoid this situation happened.In addition,the initial value of the image pose and the weight setting of the closed-loop constraint directly affect the performance and convergence speed of the SLAM back-end pose optimization process.This paper proposes a variable-weight pose image optimization algorithm.Propagation corrects the initial pose of the image that has a common view relationship with the end frame,and then sets the weight coefficient of the closed-loop constraint based on the back-propagated distance and the ratio of the matched map points of the closed-loop start-stop frame image,making the iterative convergence faster and the optimization accuracy more high.Finally,the algorithm proposed in this paper is applied to the ORB-SLAM2 system,and experiments are performed on the TUM image library and the KITTI image library.Compared with other visual SLAM systems in recent years,the trajectory error of the visual SLAM system designed in this paper smaller. |
