Loop Closure Detection

Frontend: Feature extraction, trajectory/map initial value provision
Backend: data optimization
Generated errors are accumulated by considering the relationship between adjacent frames.(인접 frame들의 관계만 고려하면 생성된 에러들이 누적됨)
- Inaccurate results in long-term estimation, overall inconsistent results
Backend-post-error estimation, but…
Good models + Bad data = Bad analysis
Loop closure detection: provide long-term constraint
- e.g., Transform the pose between x1-x100 nodes in order
- Camera passes through the same location and collects similar data
- How can you effectively detect a camera passing through the same place?
Loop Closure Detection: a very important part of the SLAM system
- Better closure detection → Better input to backend pose graph → better output
- Influences the accuracy of the map over time with estimated trajectories
- Improved accuracy and robustness of the entire SLAM
VO: front + backend, local consistency
V-SLAM: VO + loop closure/global backend, global consistency

How Loop Closure Detection Works?

Easiest way: any two images → feature correspondence → are they similar?
Any two images can have a loop back → O(n^2)
Random sample: randomly extract past data and detect loopback
As data increases, the probability determined by the loop decreases → detection efficiency decreases
Odometry based
- Detect whether there is a loopback relationship when the current camera moves to a position near the previous position
Appearance based
- Determine loop closure detection relationship based on the similarity between two images
- Similarity of images?