Method for stereo visual odometry using points, lines and planes

Patent metadata

Field	Value
Publication number	US-9430847-B2
Application number	US-201414302623-A
Country	US
Kind code	B2
Filing date	Jun 12, 2014
Priority date	Jun 12, 2014
Publication date	Aug 30, 2016
Grant date	Aug 30, 2016

Abstract

A method determines a motion between a first and second coordinate system, by first extracting a first set of primitives from a 3D image acquired in the first coordinate system from an environment, and extracting a second set of primitives from a 3D image acquired in the second coordinate system from the environment. Motion hypotheses are generated for different combinations of the first and second sets of primitives using a RANdom SAmple Consensus procedure. Each motion hypothesis is scored using a scoring function learned using parameter learning techniques. Then, a best motion hypothesis is selected as the motion between the first and second coordinate system.

First claim

We claim:

1. A method for determining a motion between a first and second coordinate system, comprising the steps of:
extracting a first set of primitives from a 3D image acquired in the first coordinate system from an environment;
extracting a second set of primitives from a 3D image acquired in the second coordinate system from the environment;
generating motion hypotheses for the first and second sets of primitives using a RANdom SAmple Consensus procedure;
determining a score for each motion hypothesis using a scoring function learned using parameter learning techniques, wherein parameters of the scoring function are determined using training data in a structured learning procedure; and
selecting, based on the score, a best motion hypothesis as the motion between the first and second coordinate system.

2. The method of claim 1 , wherein the set of primitives corresponds to points, lines and planes.

3. The method of claim 1 , wherein the points and lines are in two dimensional or three dimensional space, and the planes are in three dimensional space.

4. The method of claim 1 , wherein a subset of the primitives selected in the first coordinate system is same as the subset of the primitives selected in the second coordinate system.

5. The method of claim 1 , wherein the subset of the primitives selected in the first coordinate system is different from the subset of primitives selected in the second coordinate system.

6. The method of claim 1 , wherein the motion hypotheses are generated using combinations of the primitives as shown in Table 1.

7. The method of claim 1 , wherein the motion hypotheses are generated using 3 points in 2D space and 1 point in 3D space from the first coordinate system, and 3 points in 2D space and 1 point in 3D space from the second coordinate system.

8. The method of claim 1 , wherein the motion hypotheses are generated using 1 point in 2D space and 2 points in 3D space from the first coordinate system, and 1 point in 2D space and 2 points in 3D space from the second coordinate system.

9. The method of claim 1 , wherein a ground truth motion for the training data is obtained using a global positioning system.

10. The method of claim 1 , wherein a ground truth motion for the training data is obtained using inertial measurement units.

11. The method of claim 1 , wherein the parameters learned using the structured learning procedure are different for points, lines and planes.

12. The method of claim 1 , wherein the parameters learned using the structured learning procedure are different for different locations of the image.

13. The method of claim 1 , wherein the parameters are learned for the motion.

14. The method of claim 1 , wherein the 3D image is obtained using a stereo or trinocular stereo camera.

15. The method of claim 1 , wherein the 3D image is obtained using a red, green, blue and depth sensor.

Assignees

• Mitsubishi Electric Res Laboratories Inc

Inventors

• Ramalingam Srikumar
• Taguchi Yuichi
• Antunes Michel
• Urtasun Raquel

Classifications

• G06T7/246 Primary

  using feature-based methods, e.g. the tracking of corners or segments · CPC title

• G06T7/20 Primary

  Analysis of motion (motion estimation for coding, decoding, compressing or decompressing digital video signals H04N19/43, H04N19/51) · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06K9/52

  Physics · mapped topic

• G06T7/2033

  Physics · mapped topic