msckf_vio/pseudocode/pseudocode_image_processing

98 lines
2.4 KiB
Plaintext
Raw Normal View History

2019-04-10 18:43:30 +02:00
stereo callback()
create image pyramids
_Constructs the image pyramid which can be passed to calcOpticalFlowPyrLK._
.
if first Frame:
*initialize first Frame ()
else:
*track Features ()
*addnewFeatures ()
*pruneGridFeatures()
_removes worst features from any overflowing grid_
publish features (u1, v1, u2, v2)
_undistorts them beforehand_
addnewFeatures()
*mask existing features
*detect new fast features
*collect in a grid, keep only best n per grid
*stereomatch()
*save inliers into a new feature with u,v on cam0 and cam1
track Features()
*integrateIMUData ()
_uses existing IMU data between two frames to calc. rotation between the two frames_
*predictFeatureTracking()
_compensates the rotation between consecutive frames - rotates previous camera frame features to current camera rotation_
*calcOpticalFlowPyrLK()
_measures the change between the features in the previous frame and in the current frame (using the predicted features)_
*remove points outside of image region
_how does this even happen?_
*stereo match()
_find tracked features from optical flow in the camera 1 image_
_remove all features that could not be matched_
*twoPointRansac(cam0)
*twoPointRansac(cam1)
_remove any features outside best found ransac model_
twoPointRansac()
*mark all points as inliers
*compensate rotation between frames
*normalize points
*calculate difference bewteen previous and current points
*mark large distances (over 50 pixels currently)
*calculate mean points distance
*return if inliers (non marked) < 3
*return if motion smaller than norm pixel unit
*ransac
*optimize with found inlier after random sample
*set inlier markers
initialize first Frame()
features = FastFeatureDetector detect ()
*stereo match ()
group features into grid
- according to position in the image
- sorting them by response
- save the top responses
- save the top responses
stereo match ()
*undistort cam0 Points
*project cam0 Points to cam1 to initialize points in cam1
*calculate lk optical flow
_used because camera calibrations not perfect enough_
_also, calculation more efficient, because LK calculated anyway_
*compute relative trans/rot between cam0 and cam1*
*remove outliers based on essential matrix
_essential matrix relates points in stereo image (pinhole model)_
for every point:
- calculate epipolar line of point in cam0
- calculate error of cam1 to epipolar line
- remove if to big