search

raulmur / ORB_SLAM2

Real-Time SLAM for Monocular, Stereo and RGB-D Cameras, with Loop Detection and Relocalization Capabilities
Other
9.33k stars 4.69k forks source link

Lost tracking when large view-point changes #304

Open arenas7307979 opened 7 years ago

arenas7307979 commented 7 years ago

hi, Sometimes, when large view-point changes , the system will lost tracking (DBOW2 cannot find matching pairs)

I thinking this reason is that the Orientation by Intensity Centroid may not robust to assign the dominant orientation if i change methods (ex: sGLOH, MROGH ) to estimate the dominant orientation, will be improved?

AlejandroSilvestri commented 7 years ago

ORB-SLAM take its name from ORB (ORiented Brief), which uses BRIEF descriptors. DBOW2 uses these descriptors too.

Tracking relies on descriptor matching, in this case ORB descriptors, and doesn't use DBOW2. Changing descriptors involve a lot of coding and creating a new DBOW2 vocabulary for them. You will be doing a new visual SLAM.

On top of that, changing descriptors won't improve tracking large changes. May be motion model is not correctly predicting your large changes, thus tracking is lost.

ORB was chosen because it's fast. AKAZE is excellent, but no so fast.