Stereo Visual Odometry with Automatic Brightness Adjustment and Feature Tracking Prediction

Author: Yin

Year: 2022

Notes:

Classic indirec VO alhorithm with brightness adjustement (using an histogram matcher to predict the exposure of the image) + reliable feature point prediction (using constant acceleration model + stereo geometric constraint) to improve KLT tracker
Marginalization also (maybe the first example of vision only marg)

PL SLAM = Orbslam2 with line features
Geometric Corespondence Network (GCN) = CNN that produces point correspondence
Joint radiometric calibration and feature tracking system with an application to stereo = example of exposure estimation

brightness adjustement, feature tracking, motion estimation
Brightness adj :
- grayscale histogramm doesn't change that much at 10-20 Hz
- use instead the cumulative histogram $h$ (i.e. the probability that a pixel value is less than or equal to the given value)
- 3 parameters to evaluate the brightness: $a_{mean}, a_{half}$ and $a_{double}$ with $h(a_{mean}) = 0.5$ and so on
- Compute a coeff based on these param to evaluate if the brightness needs to be adjusted
- the comparison is done on the 20 previous frames
feature position prediction:
- ORB SLAM uses a constant velocity assumption to guide the search
- Here uses constant acceleration instead (better for dynamic motion like drones)
- simplification are made here: the increment in translation are equals, the axis of the rotation is always the same and the increments in angle of rotation are equals
- with this: predict the current pose, project all the landmarks in the current image and init a KLT tracker
- Outlier rejection using essential matrix prediction
motion prediction is perform with single frame BA with huber loss to reject outliers