Experiments On Visual Loop Closing Using Vocabulary Trees

Main idea: Several vocabulary trees as GPS data are used to deal with the problem of visual loop closing for long trajectories in an urban environment. The loop closing is performed by searching in the tree the most similar frame to the current one. Epipolar geometry is used to prune out the bad matches.

To deal with the problem of weak definitions of loop-closing a latency test is performed. This consist on testing only images that have not been taken in the last n seconds but that have been taken before that from a viewpoint in a threshold radius. Only the matched images that hold the geometric constraint are considered.

Two types of trees are tested in this work:

• Hierarchical K-means
• Extremely Randomized Forests (ERT)
  

Summarizing this approach:

1. A vocabulary tree is built off-line
2. Each image of the sequence is added to the vocabulary tree using inverted files.
3. When the vehicle is close to a previously seen scene, the descriptors of the current image and the inverted files are used to obtain the N closest images.
4. Epipolar geometry is used to determine if the matches are good.