Authors: Andrew E. Johnson, Steven B. Goldberg, Yang Cheng and Larry H. Matthies
Year: 2008
- ICRA paper ie. contribution to CV community
Considerations about MER VO:
- MER VO uses spatial corelation with Pseudo Normalized Cross Corelation (PNC) + initial motion with gyro and odometers to match features between two frames (ie no descriptor)
- MER VO uses PNC as well for stereo matching
- Use an image pyramid strategy for PNC
To improve:
- Corelation search leads to large search windows
- Too many features are rejected
MSL VO:
- Same stages as MER
- Pass through every stages at each level of the pyramid
- The bounding box for corelation search are smaller and smaller while going to the bottom of the pyramid: At each step, the bouding box for searching the matching feature is build on the previous motion estimation and the uncertainties on translation and rotation
- The covariance of the features is smaller as we go down the pyramid so that the search window becomes really small at the bottom
- Motion estimate with the left image only
- It works even with no prior on motion (eg if IMU and odometers crashes) by using the entire image at the top of the pyramid
- For the second stereo matching, the motion prior is used to define a search window on the right image
- LMS and ML as in MER motion estimator + tips for better numerical conditionning
- Validity of a transformation by checking the eigen values of the scatter matrix A (??)
Performances way better than MER VO: faster and track more features
Implemented fully in C++
Plein de petites astuces incroyables, papier très technique mais compréhensible en se concentrant bien