| As a key issue and research hotspot in the field of machine vision,Simultaneous Localization and Mapping(SLAM)is a technology that uses sensors aided robots to obtain environmental information of surrounding location and implement pose estimation as well as mapping construction.To further tackle the key and difficult issues of insufficient localization accuracy and tracking loss in SLAM,this thesis mainly introduces an adaptive robust kernel module,a multi-sensor fusion method,and a multi map fusion strategy for pose estimation optimization.In addition,this work also conducts deep research on the pose estimation and map construction parts of SLAM.The main research content of this article includes:(1)To address the problem of outlier disposal strategy under mismatched scenarios,the main research work of this thesis includes: 1)an adaptive robust kernel module for pose estimation optimization is proposed and applied to the bundle adjustment module to minimize the impact of mismatching errors on tracking process and boost the localization precision;2)via the usage of improving modules,such as the minimization of the geometric bias term,the adaptive robust kernel and the enhanced relative pose regulations,the loop candidates are jointly optimized.The experimental results illustrate that the adaptive robust kernel can effectively reduce pose estimation errors under mismatching scenarios.It ought to be emphasized that the effect of aforementioned 3 modules is much more significant in sequences with detected loop;By comparing the performance of the proposed method with other state-of-the-art SLAM algorithms,the favorable performances of the provided algorithm is evaluated.(2)In terms of the low robustness issue of SLAM approaches based on feature point method caused by feature dropout circumstances,such as,occlusion,image frame damage,and low texture features,the main research content of this work includes: 1)suggests a monocular visual inertial SLAM algorithm based on multi map fusion strategy;2)a multisensor fusion method of monocular camera and inertial measurement unit is studied,and an improved multi-map fusion strategy is also used to improve the localization accuracy and map reconstruction integrity.In addition,the performance verification on the TUM RGB-D benchmark dataset and the KITTI dataset implies that the proposed multi-map fusion algorithm is more suitable for tracking loss and mapping failure in SLAM environments due to occlusion and poor image geometry conditions.In general,the above research indicates that the proposed monocular visual SLAM method based on adaptive robust kernel BA optimization and monocular visual inertial SLAM method based on multi-map fusion strategy solve the problem of feature mismatch in the sparse maps,the low robustness scenarios caused by scene localization and the mapping under missing features such as occlusion,image frame damage,and low texture in static environments,Therefore,this article will focus on deep learning dense map construction methods in dynamic environments in the further researches. |
