Research On High Performance Loop-closure Detection Method Of Visual SLAM For Mobile Robot

Simultaneous Localization and Mapping is the core technology for mobile robots to autonomously locate and construct spatial environment maps,and it is the key to implement intelligent navigation.As a key link of visual SLAM,loop-closure detection helps to build environment maps and reduce errors of front-end estimation.However,the performance of existing loop-closure detection methods is restricted by changes in lighting and scene switching.The high-performance loop-closure detection method is the most difficult point of visual SLAM research.Therefore,we propose a visual SLAM loop-closure detection method with illumination invariance for lighting changes and proposes a cross-scene loop-closure detection method for multi-scene switching.The main research contents and results are as follows:(1)The loop-closure detection of visual SLAM under varying lighting conditions was studied.We propose a new image descriptor with illumination invariance is,which improves the adaptability of the visual SLAM loop-closure detection method under changing lighting conditions and enhances the robustness of mobile robot navigation in natural environments.Meanwhile,we propose an effective robust loop-closure detection method that fuses the original feature package.It makes image descriptors vector quantization and avoids false positives in loop-closure detection.We conducted experiments on the public datasets,which verified that the method proposed in this paper can adapt well to changes in ambient lighting and is significantly better than the Bo W method based on SIFT hand-crafted local features and the GIST method based on global features.(2)The catastrophic forgetting problem encountered by mobile robots in different scenes is studied.We propose a new loop-closure detection method with continual learning capabilities,which improves the learning process and network structure of CNN and enhances the cross-scene learning capabilities of visual SLAM.The method enables the loop-closure detection to improve the accuracy of the robot’s recognition of the scenes while maintaining the ability to accurately recognize the old scene when learning scenarios increases.It makes up for the shortcomings of the continual learning ability of the robots SLAM system and lays a theoretical and technical foundation for further promoting the research of human SLAM.At the same time,a set of solutions to evaluate the performance of visual SLAM is proposed,which can evaluate the crossscene performance of the latest research results of visual SLAM.We verified the performance of our method on public datasets,and the experiments show that our method is superior to Net VLAD which is the representative loop-closure detection method.