Consistent Ground-Plane Mapping: A Case Study Utilizing Low-Cost Sensor Measurements and a Satellite Image

Goal

Create a high resolution roadway image for autonomous driving, only using low-cost conventional sensors. The created image should be consistent with the a prior satellite image.

Challenge

Low-cost sensors cannot localize accurately, must resort to matching input and satellite images to achieve consistency, but matching does not always work.
Example: For the first and second images, I have something to match with. What about the third one? It could (and actually does) show strong correlation at random wrong location.

roadway img1 roadway img2 roadway img3

Method

Step 1: Classify anchor frames: frames that have highly identifiable structured textures.
Example: In the above figure, first and second are anchor frames, third is not.
Step 2: Adjacent anchor frames form an anchor, make adjustment in anchor units.
Example: This not only avoids "bad" images, but also produces smooth adjustments.

anchor-wise method
Finally we got this:

icra15_demo_HChu_AVu from Hang Chu on Vimeo.

References

[1] Hang Chu, and Anh Vu, Consistent Ground-Plane Mapping: A Case Study Utilizing Low-Cost Sensor Measurements and a Satellite Image, submitted to IEEE International Conference on Robotics and Automation (ICRA), 2015. [pdf]