Research On Robust Visual Place Recognition Method For Mobile Robots In Dramatic Illumination Environment

Simultaneous Localization and Mapping is the key technology to realize autonomous navigation of mobile robots.As an important part of visual SLAM,Visual Place Recognition can realize robot location.At the same time,the robustness of visual place recognition to changes in environmental conditions affects the accuracy of robots in constructing environmental maps.Therefore,how to improve the robustness of visual place recognition is one of the key scientific issues in the research direction of visual SLAM.However,when the robot moves in the outdoor natural environment for a long time,it will inevitably face the challenge of drastic illumination changes(time shift,season,or rain and fog),which causes the robot's place recognition ability to be greatly restricted.This paper proposes a method for robot visual place recognition that is more robust to severe illumination changes.Firstly,GAN is introduced into visual SLAM to enhance the illumination quality of robot map candidate keyframe images,which provides a guarantee for the subsequent extraction of rich invariant features.Secondly,this paper proposes a method for evaluating the geographic location authenticity with the enhanced quality of map candidate keyframes,based on the global image descriptor to ensure that subsequent image matches correspond to the correct geographic location.Next,a deep convolutional neural network is used to extract the illumination invariant feature DELF from the map candidate keyframe image with enhanced illumination quality and an environment invariant image descriptor based on the DELF feature vector matrix is generated for image matching.Finally,a robust visual place recognition of robot with illumination changes is realized.On the Oxford Robot Car,a public dataset containing severely changing illumination environments,experiments verify the effectiveness of the proposed method for evaluating the true geographic location expression after the map candidate keyframes are enhanced in quality,and then verify the invariance of the environmental descriptors for the image descriptors proposed in this paper.Then,the robustness of the visual place recognition method proposed in this paper is verified,and the method in this paper is compared with the existing representative methods.The experimental results show that the method in this paper is more significant than the existing methods in terms of robustness.The visual place recognition method proposed in this paper enables the robot to show a highly robust place recognition ability in the face of changing environments with severe illumination,which is conducive to improving the quality of keyframes' node in the environmental map,achieving accurate location of the robot,providing strong guarantees to autonomous navigation.