Relocalization module is meant to localize a robot given a previously gathered map using 1 or more images.
The two main methods of relocalization are:
Image2Map: which checks for feature correspondences in the query image directly to the 3d map
Image2Image: which checks for the best matching image in the dataset and localizes wrt to it
However our case of relocalization is quite unique due to our access to different pieces of data. We can assume we not only have access to the raw images of a map, but also the triangulated location of features in the images, and labelled defects.
The solution proposed would therefore be a hybrid:
Determine the best matching image using defect enhanced bag of words
get initial estimate using ransac-8-point algorithm
pose refinement using co-visibility of features and minimize reprojection error
Relocalization module is meant to localize a robot given a previously gathered map using 1 or more images.
The two main methods of relocalization are:
However our case of relocalization is quite unique due to our access to different pieces of data. We can assume we not only have access to the raw images of a map, but also the triangulated location of features in the images, and labelled defects.
The solution proposed would therefore be a hybrid: