| Simultaneous Localization and Mapping(SLAM)is the key technology for mobile robot to realize autonomous navigation.Determining the position of a device(such as a robot,camera,or vision sensor,etc.)which is in an unfamiliar environment quickly and accurately is a core issue to improve the quality of the map.Currently,with the rapid development of virtual reality,driverless,and 3D reconstruction technologies,a single sensor which can only acquire a small amount of information has been unable to meet the requirement of the reconstruction of wide scene,and it usually requires multiple sensors to capture images from different angles.Therefore,the loop closure detection method was studied by using multi-view fusion data in this thesis.Matterport 3D Pro camera is used to scan the real scene to obtain RGB-D image data.Firstly,the acquired color images were fused to obtain a sequence of panoramic images,and the SURF feature of each image was extracted.Secondly,depth information was added to each feature point according to the depth map corresponding to the color image,and the extracted image features were preprocessed.Then,a visual dictionary tree was constructed based on the pre-processed features combining the visual dictionary technology in computer vision.In addition,the proposed method above was verified using the public RGB-D data set.Since the visual data in the dataset was sensitive to light conditions,it was difficult to extract valid features in the visually weak areas.Therefore,the multi-frame depth map fusion 3D point cloud data was used to train visual dictionary.Finally,in order to ensure the reliability of the closed-loop detection,point cloud data was registered using the RANSAC and ICP algorithm,and its overlapping area was treated as a standard of geometric verification.Based on the above two kinds of data sets,the closed-loop detection method proposed in this paper was tested.The experimental results showed that compared with the method of any two image matching,the proposed one avoids the matching complexity of O(N~2),had good robustness and accuracy,and could realize automatic global registration of 3D model. |
