Image-based localization

• US 8,798,357 B2
• Filed: 07/09/2012
• Issued: 08/05/2014
• Est. Priority Date: 07/09/2012
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Sinha, Sudipta N, Lim, Hyon, Cohen, Michael, Uyttendaele, Matt
• Primary Examiner(s): BHATNAGAR, ANAND P
• Application Number:
  US13/544,817
• Publication Number:
  US 20140010407A1
• Time in Patent Office: 757 Days
• Field of Search: 382/100, 382/103, 382/154, 382/168, 382/170, 345419-427, 356/12
• US Class Current: 382/154
• CPC Class Codes: G06T 2207/10016 Video; Image sequence
• CPC Class Codes: G06T 2207/20016 Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
• CPC Class Codes: G06T 2207/30244 Camera pose
• CPC Class Codes: G06T 7/73 using feature-based methods
• CPC Class Codes: G06V 20/10 Terrestrial scenes scenes under surveillance with static cameras G06V20/52; scenes perceived from the exterior of a vehicle G06V20/56; scenes perceived from the interior of a vehicle G06V20/59

Abstract

Image-based localization technique embodiments are presented which provide a real-time approach for image-based video camera localization within large scenes that have been reconstructed offline using structure from motion or similar techniques. From monocular video, a precise 3D position and 3D orientation of the camera can be estimated on a frame by frame basis using only visual features.

Claims

1. A computer-implemented process for computing the 3D position and 3D orientation of a video camera used to capture sequential image frames of an environment for which a three dimensional (3D) reconstruction has been pre-computed to identify 3D points in the environment that correspond to two dimensional (2D) points in previously-captured images of the environment, comprising:

using a computer to perform the following process actions;

constructing an indexed database comprising multiple representative database descriptors for each 3D point in the 3D reconstruction, ones of said database descriptors for each 3D point being computed at different scales and from multiple ones of said previously-captured images of the environment;

after the indexed database is constructed, inputting image frames from the video camera as they are captured;

for each consecutive image frame input,

identifying tracked keypoints representing 2D image frame locations in the image frame, said tracked keypoints being either newly added or previously depicted and identified in one or more previously-input image frames,

computing a database descriptor for each newly added tracked keypoint identified in the image frame, said database descriptor being of the same type as computed for the indexed database,

identifying 3D points in the 3D reconstruction of the environment that correspond to the tracked keypoints, said identifying of 3D points comprising, for each tracked keypoint, matching the database descriptor computed for the tracked keypoint to one or more descriptors in the indexed database and determining the 3D point associated with the matched database descriptors in the indexed database, and

estimating the 3D position and 3D orientation of the video camera.

2. The process of claim 1, wherein the process action of constructing the indexed database, comprises the actions of:

for each of said previously-captured images of the environment,

generating an image pyramid comprising a plurality of levels each of which depicts the scene captured by the image at a different scale;

for each level of the image pyramid of the image,

identifying keypoints,

computing a database descriptor for each identified keypoint,

generating a database comprising each database descriptor as an entry and assigning to each database descriptor a set of indices indicating which previously-captured image of the environment the keypoint associated with the descriptor came from and the image coordinates of that keypoint; and

constructing a k-d tree representation index for the database descriptors in the database.

3. The process of claim 2, wherein the process action of constructing the indexed database, further comprises an action of constructing a look-up table that maps each database descriptor in the database to a 3D point in the 3D reconstruction.

4. The process of claim 3, wherein the process action of constructing the look-up table, comprises the actions of:

for each level of the image pyramid of each previously-captured image of the environment,

projecting each 3D point from the 3D reconstruction, which is depicted in the image under consideration, into the pyramid level of the image, and

for each keypoint previously identified in the pyramid level of the image,

identifying the projected 3D point that is closest to the keypoint,

determining if the identified closest projected 3D point is within a threshold distance of the keypoint, and

whenever it is determined that the identified closest projected 3D point is within the threshold distance, assigning the keypoint and its associated database descriptor to the 3D point corresponding to the closest projected 3D point; and

creating an entry in said look-up table listing each database descriptor in the indexed database along with the 3D point in the 3D reconstruction assigned thereto and the keypoint associated with the database descriptor.

5. The process of claim 4, wherein the process action of constructing the look-up table, further comprises the actions of:

grouping the previously-captured images of the environment used to produce the 3D reconstruction into overlapping clusters based on the 3D position and 3D orientation of the camera used to capture each of the previously-captured images;

assigning a unique group identifier to each cluster of images;

for each cluster of images,

identifying the database descriptors associated with each image in the cluster using said indexed database, and

adding the group identifier assigned to the cluster to the entry in said look-up table pertaining to each of the identified database descriptors.

6. The process of claim 5, wherein the process action of identifying keypoints in the currently input image frame which were depicted and identified in one or more previously-input image frames, comprises the actions of:

using a Harris corner detection procedure to identify candidate keypoints in the current image frame;

for each candidate keypoint, determining whether a Harris corner cornerness value computed for the candidate keypoint exceeds an adaptive cornerness threshold, said adaptive cornerness threshold being computed as a prescribed percentage of the maximum cornerness value in the last, previously-input image frame;

whenever a candidate keypoint is found to have a Harris corner cornerness value that exceeds the adaptive cornerness threshold or for all the candidate keypoints if the current image frame is the first-input image frame, designating the candidate keypoint as a keypoint in the current image frame, computing a Binary Robust Independent Elementary Feature (BRIEF) descriptor for a prescribed sized and shaped pixel patch centered around the keypoint and assigning the BRIEF descriptor to the keypoint;

for each keypoint designated as a tracked keypoint in the last, previously-input image frame,

establishing a prescribed sized and shaped search window that is centered on a location in the current image frame that corresponds to the coordinates of the tracked keypoint under consideration in the last, previously-input image frame,

computing a difference measure between each BRIEF descriptors associated with the keypoints in the current image frame that lie within said established search window and a BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame,

determining which BRIEF descriptor associated with the keypoints in the current image frame that lie within said established search window is the most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame based on the computed difference measures,

designating the keypoint in the current input image frame whose BRIEF descriptor was found to be most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame to be a tracked keypoint which corresponds to the tracked keypoint that was depicted and identified in one or more previously-input image frames, and which is associated with the same database descriptor as was computed for the tracked keypoint that was depicted and identified in one or more previously-input image frames.

7. The process of claim 6, wherein the process action of computing difference measures, comprises an action of computing a Hamming distance between each BRIEF descriptors associated with the keypoints in the current image frame that lie within said established search window and a BRIEF descriptor computed for the tracked keypoint under consideration from the last, previously-input image frame.

8. The process of claim 6, further comprising performing, prior to performing the process action of computing a database descriptor for each newly added tracked keypoint identified in the image frame, the actions of:

generating a current version of a tracking table, said tracking table comprising a listing of tracked keypoints along with, for each keypoint listed, a BRIEF descriptor computed for the keypoint, and if known, a 2D-3D point correspondence which identifies the coordinates of a 3D point in the 3D reconstruction that corresponds the 2D tracked keypoint;

determining if the current version of the tracking table indicates that the number of tracked keypoints listed exceeds a prescribed minimum tracked keypoint count;

whenever the tracked keypoint count does not exceed the prescribed minimum tracked keypoint count,

identifying areas in the current image frame not having tracked keypoints therein,

for each identified area, adding keypoints identified in the current image frame that lie in the identified area, but which are not designated as tracked keypoints, to the current version of the track table, along with the previously-computed BRIEF descriptors assigned thereto, and designating the added keypoints to be tracked keypoints, to produce a new current version of the tracking table.

9. The process of claim 8, further comprising:

performing, prior to performing the process action of estimating the 3D position and 3D orientation of the video camera, the actions of,

tallying the number of 2D-3D point correspondences listed in the current version of the tracking table,

determining if the tallied number exceeds a prescribed 2D-3D point correspondences threshold number;

whenever it is determined that the tallied number does not exceed the prescribed 2D-3D point correspondences threshold number,

for each database descriptor computed for a tracked keypoint identified in the current image frame,

performing a k-approximate nearest neighbor (ANN) query search of the indexed database using the k-d tree representation index to obtain a listing of nearest neighbor descriptors,

identifying a distance measure between each nearest neighbor descriptor in the listing and the database descriptor under consideration, said distance measure increasing in value the more a nearest neighbor descriptor differs from the database descriptor under consideration,

for each nearest neighbor descriptor in the listing, determining if its distance measure is less than a prescribed distance,

whenever the distance measure of a nearest neighbor descriptor in the listing is less than the prescribed distance, identifying the 3D point using said look-up table that corresponds to the nearest neighbor descriptor from the listing,

identifying each cluster of images that the identified 3D point belongs to based on the group identifier associated with the identified 3D in the look-up table,

assigning a score to each identified cluster that is equal to the prescribed distance divided by the distance measure of the nearest neighbor descriptor under consideration,

ascertaining the highest score assigned to the identified clusters,

computing a minimum score, said minimum score being defined as a prescribed percentage of the ascertained highest score,

identifying those clusters having a score that is equal to or exceeds the minimum score and identifying the database images belonging to the identified clusters,

eliminating from the listing those nearest neighbor descriptors that are not associated with the identified database images to produce a refined listing,

for each 3D point corresponding to one or more nearest neighbor descriptors in the refined listing, computing a matching strength value, said matching strength value being defined as the summation of the scores of the nearest neighbor descriptors associated with the 3D point, and

for the keypoint corresponding to the descriptor under consideration,

identifying the 3D point having the greatest matching strength value and the 3D point having the second greatest matching strength value,

determining if the second greatest matching strength value divided by the greatest matching strength value for the 3D points exceeds a prescribed ratio value,

whenever it is determined that the second greatest matching strength value divided by the greatest matching strength value exceeds the prescribed ratio value, assigning the 3D point having the greatest matching strength value to the keypoint to form a 2D-3D point correspondence,

computing an estimate of the current 3D position and 3D orientation of a video camera by applying the last-estimated translational and angular velocities to the last-estimated 3D position and 3D orientation of the video camera for the immediate preceding frame,

performing a geometric verification procedure on each newly formed 2D-3D point correspondence using the last-estimated camera pose and eliminating those newly formed 2D-3D point correspondences that do not match within a prescribed tolerance,

respectively adding each of the remaining newly formed 2D-3D point correspondences to the tracking table in the listing associated with the tracked keypoint associated with the 2D-3D point correspondence, and

designating the updated version of the tracking table to be the new current tracking table.

10. The process of claim 9, further comprising the process actions of:

whenever one or more newly-formed 2D-3D point correspondences are added to the tracking table,

tallying the number of 2D-3D point correspondences listed in the current version of the tracking table,

determining if the tallied number of 2D-3D point correspondences exceeds said prescribed 2D-3D point correspondences threshold number,

whenever it is determined the tallied number of 2D-3D point correspondences exceeds said prescribed 2D-3D point correspondences threshold number, performing said process action of estimating the 3D position and 3D orientation of the video camera using the 2D-3D correspondences between at least some of the identified 3D points and the corresponding 2D location of the tracked keypoints in the current frame, and

whenever it is determined the tallied number of 2D-3D point correspondences does not exceed said prescribed 2D-3D point correspondences threshold number, performing said process action of estimating the 3D position and 3D orientation of the video camera by designating the 3D position and 3D orientation of the video camera computed by applying the last-estimated translational and angular velocities to the last-estimated 3D position and 3D orientation of the video camera for the immediate preceding frame to be the estimate for the current image frame, and

updating the current translational and angular velocities of the video camera using a Kalman filter updating procedure.

11. The process of claim 8, further comprising the actions of:

whenever it is determined that the tallied number does exceed the prescribed 2D-3D point correspondences threshold number, performing said process action of estimating the 3D position and 3D orientation of the video camera using the 2D-3D correspondences between at least some of the identified 3D points and their corresponding 2D tracked keypoints;

determining if there are tracked keypoints listed in the most current version of the tracking table that do not have a 2D-3D point correspondence associated therewith;

whenever it is determined that there are no tracked keypoints listed in the most current version of the tracking table that do not have a 2D-3D point correspondence associated therewith, updating the current translational and angular velocities of the moving video camera using a Kalman filter updating procedure; and

whenever it is determined that there are tracked keypoints listed in the most current version of the tracking table that do not have a 2D-3D point correspondence associated therewith,

for each image associated with a database descriptor in the indexed database, tallying the number of 2D-3D correspondences listed in the most current version of the tracking table that are depicted in the image,

ranking the images in decending order based on the 2D-3D correspondences tallies,

identifying a prescribed number of the top ranking images,

for at least some of the database descriptors in the indexed database associated with a tracked keypoint depicted in the top ranking images that does not have a 2D-3D point correspondence associated therewith,

performing a k-approximate nearest neighbor (ANN) query search of the indexed database using the k-d tree representation index to obtain a listing of nearest neighbor descriptors,

identifying a distance measure between each nearest neighbor descriptor in the listing and the database descriptor under consideration, said distance measure increasing in value the more a nearest neighbor descriptor differs from the database descriptor under consideration,

for each nearest neighbor descriptor in the listing,

determining if its distance measure is less than a prescribed distance,

whenever the distance measure of a nearest neighbor descriptor in the listing is less than the prescribed distance, identifying the 3D point using said look-up table that corresponds to the nearest neighbor descriptor from the listing,

assigning a score to the nearest neighbor descriptor that is equal to the prescribed distance divided by the distance measure of the nearest neighbor descriptor,

for each identified 3D point corresponding to one or more nearest neighbor descriptors in the listing, computing a matching strength value, said matching strength value being defined as the summation of the scores of the nearest neighbor descriptors associated with the 3D point,

assigning each identified 3D point having a ratio of its matching strength value to the highest matching strength value among the identified 3D points that exceeds the prescribed ratio value to the tracked keypoint corresponding to the database descriptor under consideration to form 2D-3D point correspondences,

performing a geometric verification procedure on each newly formed 2D-3D point correspondence using the last-estimated camera pose and eliminating those newly formed 2D-3D point correspondences that do not match within a prescribed tolerance,

respectively adding each of the remaining newly formed 2D-3D point correspondences to the tracking table in the listing associated with the tracked keypoint associated with the 2D-3D point correspondence, and

designating the updated version of the tracking table to be the new current tracking table.

12. The process of claim 11, further comprising the process actions of:

whenever one or more newly-formed 2D-3D point correspondences are added to the tracking table,

performing said process action of estimating the 3D position and 3D orientation of the video camera using the 2D-3D correspondences between at least some of the identified 3D points and their corresponding 2D tracked keypoints, and

updating the current translational and angular velocities of the video camera using a Kalman filter updating procedure.

13. The process of claim 11, wherein said at least some of the database descriptors in the indexed database associated with a tracked keypoint depicted in the top ranking images that does not have a 2D-3D point correspondence associated therewith, is defined as all the database descriptors in the indexed database associated with a tracked keypoint depicted in the top ranking images that does not have a 2D-3D point correspondence associated therewith up to a prescribed maximum number.

14. The process of claim 5, wherein the process action of identifying keypoints in the currently input image frame which were depicted and identified in one or more previously-input image frames, comprises the actions of:

using a Harris corner detection procedure to identify candidate keypoints in the current image frame;

for each candidate keypoint, determining whether a Harris corner cornerness value computed for the candidate keypoint exceeds an adaptive cornerness threshold, said adaptive cornerness threshold being computed as a prescribed percentage of the maximum cornerness value in the last, previously-input image frame;

whenever a candidate keypoint is found to have a Harris corner cornerness value that exceeds the adaptive cornerness threshold or for all the candidate keypoints if the current image frame is the first-input image frame, designating the candidate keypoint as a keypoint in the current image frame, computing a Binary Robust Independent Elementary Feature (BRIEF) descriptor for a prescribed sized and shaped pixel patch centered around the keypoint and assigning the BRIEF descriptor to the keypoint;

for each keypoint designated as a tracked keypoint in the last, previously-input image frame,

establishing a prescribed sized and shaped search window that is centered on a location in the current image frame that corresponds to the coordinates of the tracked keypoint under consideration in the last, previously-input image frame,

computing a difference measure between each BRIEF descriptors associated with the keypoints in the current image frame that lie within said established search window and a BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame,

determining which BRIEF descriptors associated with the keypoints in the current image frame that lie within said established search window are the most similar and the second most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame based on the computed difference measures,

computing a ratio of the difference measure computed for the current image frame keypoint BRIEF descriptor determined to be most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame to the difference measure computed for the current image frame keypoint BRIEF descriptor representing the second most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame; and

whenever the computed ratio is less than a prescribed number, designating the current image frame keypoint whose BRIEF descriptor was determined to be most similar to the BRIEF descriptor computed for the tracked keypoint under consideration in the last, previously-input image frame as a tracked keypoint which corresponds to the tracked keypoint that was depicted and identified in one or more previously-input image frames, and which is associated with the same database descriptor as was computed for the tracked keypoint that was depicted and identified in one or more previously-input image frames.

15. The process of claim 2, wherein the process action of identifying keypoints, comprises an action of employing a Harris corner detection procedure.

16. The process of claim 2, wherein the process action of computing a database descriptor for each identified keypoint, comprises the actions of:

for each of the identified keypoints,

identifying the orientation of the keypoint associated with a peak in a gradient orientation histogram, said orientation being defined in terms of an angle of rotation of a prescribed patch axis about the keypoint from a prescribed image axis,

sampling a patch of pixels from the level of the image pyramid of the image under consideration, said patch having a prescribed shape and size which is centered on the keypoint and which is rotated such that said patch axis forms the identified orientation angle with said image axis,

computing a DAISY descriptor for the sampled patch of pixels, and

establishing a database descriptor for the keypoint comprising said orientation angle and said DAISY descriptor.

17. A computer-implemented process for computing the 3D position and 3D orientation of a video camera used to capture sequential image frames of an environment for which a three dimensional (3D) reconstruction has been pre-computed to identify 3D points in the environment that correspond to two dimensional (2D) points in previously-captured images of the environment, comprising:

using a computer to perform the following process actions;

constructing an indexed database comprising multiple representative database descriptors for each 3D point in the 3D reconstruction, ones of said database descriptors for each 3D point being computed at different scales and from multiple ones of said previously-captured images of the environment;

after the indexed database is constructed, inputting image frames from the video camera as they are captured;

for each consecutive image frame input,

identifying tracked keypoints representing 2D image frame locations in the image frame, said tracked keypoints being either newly added or previously depicted and identified in one or more previously-input image frames,

computing a database descriptor for each newly added tracked keypoint identified in the image frame, said database descriptor being of the same type as computed for the indexed database,

identifying 3D points in the 3D reconstruction of the environment that correspond to the tracked keypoints to produce 2D-3D point correspondences each of which identifies the coordinates of a 3D point in the 3D reconstruction that corresponds the 2D tracked keypoint, said identifying of 3D points comprising, for each tracked keypoint, matching the database descriptor computed for the tracked keypoint to one or more descriptors in the indexed database and determining the 3D point associated with the matched database descriptors in the indexed database,

initially estimating the 3D position and 3D orientation of the video camera using a random sample consensus (RANSAC) procedure with three-point pose estimation to identify a set of inliers among the 2D-3D point correspondences,

refining the initial estimate of the 3D position and 3D orientation of the video camera using a non-linear least squares optimization procedure to produce a current estimate of the video camera 3D position and 3D orientation for the image frame under consideration,

determining whether the number of 2D-3D point correspondences identified as inliers exceeds a prescribed correspondences number,

whenever it is determined that the number of 2D-3D point correspondences identified as inliers exceeds the prescribed correspondences number, updating the current translational and angular velocities of the video camera using a Kalman filter updating procedure.

18. A computer-implemented process for computing the 3D position and 3D orientation of a video camera used to capture sequential image frames of an environment for which a three dimensional (3D) reconstruction has been pre-computed to identify 3D points in the environment that correspond to two dimensional (2D) points in previously-captured images of the environment, comprising:

using a computer to perform the following process actions;

constructing an indexed database comprising multiple representative database descriptors for each 3D point in the 3D reconstruction, ones of said database descriptors for each 3D point being computed at different scales and from multiple ones of said previously-captured images of the environment;

after the indexed database is constructed, inputting image frames from the video camera as they are captured;

for each consecutive image frame input,

identifying tracked keypoints representing 2D image frame locations in the image frame, said tracked keypoints being either newly added or previously depicted and identified in one or more previously-input image frames,

computing a database descriptor for each newly added tracked keypoint identified in the image frame, said database descriptor being of the same type as computed for the indexed database,

identifying 3D points in the 3D reconstruction of the environment that correspond to the tracked keypoints to produce 2D-3D point correspondences each of which identifies the coordinates of a 3D point in the 3D reconstruction that corresponds the 2D tracked keypoint, said identifying of 3D points comprising, for each tracked keypoint, matching the database descriptor computed for the tracked keypoint to one or more descriptors in the indexed database and determining the 3D point associated with the matched database descriptors in the indexed database,

grouping the previously-captured images of the environment used to produce the 3D reconstruction into overlapping clusters based on the 3D position and 3D orientation of the camera used to capture each of the previously-captured images,

assigning a unique group identifier to each cluster of images;

for each cluster of images,

identifying the database descriptors associated with each image in the cluster using said indexed database,

tallying the number of 2D-3D point correspondences,

determining if the tallied number exceeds a prescribed 2D-3D point correspondences threshold number;

whenever it is determined that the tallied number does not exceed the prescribed 2D-3D point correspondences threshold number,

for each database descriptor computed for a tracked keypoint identified in the current image frame,

performing a search of the indexed database to obtain a listing of nearest neighbor descriptors,

identifying a distance measure between each nearest neighbor descriptor in the listing and the database descriptor under consideration, said difference measure increasing in value the more a nearest neighbor descriptor differs from the database descriptor under consideration,

for each nearest neighbor descriptor,

determining if its distance measure is less than a prescribed distance,

whenever the distance measure of a nearest neighbor descriptor in the listing is less than the prescribed distance, identifying the 3D point that corresponds to the nearest neighbor descriptor,

 

identifying each cluster of images that the identified 3D point belongs to,

 

assigning a score to each identified cluster that is equal to the prescribed distance divided by the distance measure of the nearest neighbor descriptor under consideration,

ascertaining the highest score assigned to the identified clusters,

computing a minimum score, said minimum score being defined as a prescribed percentage of the ascertained highest score,

identifying those clusters having a score that is equal to or exceeds the minimum score and identifying the database images belonging to the identified clusters,

eliminating those nearest neighbor descriptors that are not associated with the identified database images,

for each 3D point corresponding to one or more remaining nearest neighbor descriptors, computing a matching strength value, said matching strength value being defined as the summation of the scores of the nearest neighbor descriptors associated with the 3D point, and

for the keypoint corresponding to the descriptor under consideration,

identifying the 3D point having the greatest matching strength value and the 3D point having the second greatest matching strength value,

determining if the second greatest matching strength value divided by the greatest matching strength value for the 3D points exceeds a prescribed ratio value,

whenever it is determined that the second greatest matching strength value divided by the greatest matching strength value exceeds the prescribed ratio value, assigning the 3D point having the greatest matching strength value to the keypoint to form a 2D-3D point correspondence,

computing an estimate of the current 3D position and 3D orientation of a video camera by applying the last-estimated translational and angular velocities to the last-estimated 3D position and 3D orientation of the video camera for the immediate preceding frame,

performing a geometric verification procedure on each newly formed 2D-3D point correspondence using the last-estimated camera pose and eliminating those newly formed 2D-3D point correspondences that do not match within a prescribed tolerance,

re-tallying the number of 2D-3D point correspondences,

determining if the re-tallied number of 2D-3D point correspondences exceeds said prescribed 2D-3D point correspondences threshold number,

whenever it is determined the re-tallied number of 2D-3D point correspondences exceeds said prescribed 2D-3D point correspondences threshold number, estimating the 3D position and 3D orientation of the video camera using the 2D-3D correspondences between at least some of the identified 3D points and their corresponding 2D tracked keypoints, and

estimating the current translational and angular velocities of the 3D position and 3D orientation of a video camera based on the current estimate of the 3D position and 3D orientation of the video camera.

19. The process of claim 18 further comprising the process actions of:

whenever it is determined the re-tallied number of 2D-3D point correspondences does not exceed said prescribed 2D-3D point correspondences threshold number, designating the 3D position and 3D orientation of the video camera computed by applying the last-estimated translational and angular velocities to the last-estimated 3D position and 3D orientation of the video camera for the immediate preceding frame to be the estimate for the current image frame; and

estimating the current translational and angular velocities of the 3D position and 3D orientation of a video camera based on the current estimate of the 3D position and 3D orientation of the video camera.

20. The process of claim 18, wherein the process action of constructing the indexed database further comprising the actions of:

for each of said previously-captured images of the environment,

generating an image pyramid comprising a plurality of levels each of which depicts the scene captured by the image at a different scale;

for each level of the image pyramid of the image,

identifying keypoints,

computing a database descriptor for each identified keypoint,

generating a database comprising each database descriptor as an entry and assigning to each database descriptor a set of indices indicating which previously-captured image of the environment the keypoint associated with the descriptor came from and the image coordinates of that keypoint; and

constructing a k-d tree representation index for the database descriptors in the database; and

wherein

the process action of performing the search of the indexed database to obtain the listing of nearest neighbor descriptors, comprises an action of performing a k-approximate nearest neighbor (ANN) query search of the indexed database using the k-d tree representation index to obtain a listing of nearest neighbor descriptors.