Monocular camera localization in large scale indoor sparse LiDAR point cloud
A method of camera localization, comprising, receiving a camera image, receiving a LiDAR point cloud, estimating an initial camera pose for the camera image, sampling an initial set of depth projections within the LiDAR point cloud, measuring a similarity of the initial camera pose to the initial set of depth projections and deriving a subsequent set of depth projections based on the measured similarity.
1. A method of camera localization, comprising:
receiving a camera image;
receiving a LiDAR point cloud;
estimating an initial camera pose for the camera image;
sampling an initial set of depth projections within the LiDAR point cloud;
measuring a similarity of the initial camera pose to the initial set of depth projections; and
deriving a subsequent set of depth projections based on the measured similarity in a camera image feature space.
2. The method of camera localization of claim 1 wherein the deriving a subsequent set of depth pose projections is based on minimization of the measured similarity.
3. The method of camera localization of claim 1 further comprising back-propagating the derivation of the subsequent set of depth projections.
4. The method of camera localization of claim 1 wherein the measured similarity is a state-value based function.
5. The method of camera localization of claim 1 further comprising measuring a pose loss.
6. The method of camera localization of claim 1 wherein the LiDAR point cloud is a sparse dataset.
7. The method of camera localization of claim 1 further comprising subsequently deriving the subsequent set of depth projections based on the measured similarity until convergence.
8. The method of camera localization of claim 4 further comprising regressing a ground truth pose difference based on the measured similarity.
9. The method of camera localization of claim 8 further comprising back-propagating the state-value based function as a gradient and the ground truth pose difference.
10. A method of camera localization, comprising:
receiving a camera image;
receiving a LiDAR point cloud;
estimating an initial camera pose for the camera image;
sampling an initial set of depth projections within the LiDAR point cloud;
measuring a similarity of the initial camera pose to the initial set of depth projections as a state value;
measuring a regression of the similarity of the initial camera pose to the initial set of depth projections;
measuring a gradient of the similarity of the initial camera pose to the initial set of depth projections; and
deriving a subsequent set of depth projections based on the gradient of the similarity in a camera image feature space.
11. The method of camera localization of claim 10 wherein the measuring of the similarity is performed by a first network.
12. The method of camera localization of claim 11 wherein the measuring of the regression of the similarity is performed by a second network.
13. The method of camera localization of claim 12 further comprising weight sharing between the first network and the second network.
14. The method of camera localization of claim 13 further comprising restraining the state value of the measured similarity fed back to the first network and the second network.
15. The method of camera localization of claim 14 wherein the restraining the state value of the measured similarity fed back to the first network and the second network is performed by a resistor module.
16. The method of camera localization of claim 15 further comprising subsequently deriving the subsequent set of depth projections based on the measured similarity until convergence.