Research On Closed Loop Detection Algorithm Based On Pose Constraint In Visual SLAM

With the wide application of mobile robots in people's daily life,visual SLAM technology has become one of the key technologies,and has gradually become a research hotspot in recent years.Vision-based closed-loop detection is an important part of visual SLAM.The purpose is to identify the places that the robot has visited before,so as to correct the accumulating pose errors during driving and obtain more accurate robot pose estimation.Simply relying on visual information to achieve closed-loop detection,in the similar scene or repeated appearance environment,it is easy to produce perceptual ambiguity,which leads to mismatching;in the case of large changes in ambient illumination,it is easy to miss the closed-loop location.Aiming at these two problems,the proposed algorithm adopts a kind of descriptor that is robust to illumination to realize closed-loop detection,which solves the problem of closed-loop loss under large illumination changes;combines the attitude information of the odometer with the visual closed-loop detection process.A closed-loop visual closed-loop detection method is proposed to solve the problem of perceived ambiguity.The paper is mainly introduced from three aspects.Firstly,the global and local descriptors are introduced for the instability of the features under illumination variation conditions.The illumination robust global descriptor DIRD is described in detail,including image filtering,normalization and quantization.The construction of the descriptor.Then,the calculation method of similarity between images is introduced and the composition of the descriptor is shown by experiments.Secondly,the method of camera pose estimation and the error estimation process are introduced,which provides the constrained pose for the next visual closed-loop detection.In this paper,the MSCKF(Multi-State Constraint Kalman Filter)algorithm based on filtering method is used to realize the camera pose estimation of image and INS information fusion.The process of its implementation is described and experiments are carried out on the actual data set.The estimated position of the algorithm is analyzed.Error and attitude errors lay the foundation for the implementation of the next pose constraint algorithm.Thirdly,the relationship between visual closed-loop detection and pose constraint is systematically studied.A closed-loop detection system framework of DIRD under pose constraints is proposed.The algorithm uses the provided pose constraint to determine the closed-loop candidate region,uses the DIRD operator to achieve closed-loop detection,and eliminates the mismatch caused by the perceived ambiguity according to the position constraint.The algorithm is validated on the KITTI dataset.The Precision-Recall curve is used as the evaluation standard.The accuracy and real-time performance of the closed-loop detection algorithm are evaluated and analyzed.The experimental results show that the closed-loop detection method proposed in this paper is better.The feasibility and effectiveness not only improve the efficiency and accuracy of closed-loop detection,but also have good robustness.