Place Recognition

To close a loop, a robot must notice that it has been here before. Place recognition compares what the robot sees now with a database of past views. A classic method describes each view as a bag of visual words, a histogram of the kinds of features it contains.

Robot is at place Viewpoint and lighting change Two places look almost the same Match threshold

Bag of words

Features in each image are assigned to the nearest of $K$ visual words, learned in advance. A view becomes a weighted histogram $\mathbf{v}$, and two views are compared by cosine similarity:

$$s(\mathbf{a}, \mathbf{b}) = \frac{\mathbf{a}\cdot\mathbf{b}}{\lVert\mathbf{a}\rVert\,\lVert\mathbf{b}\rVert}$$

Words that appear everywhere are down-weighted, which is the tf-idf idea from text search. DBoW2, used in ORB-SLAM, works this way.

What to notice

The true place scores highest, and the bars for other places stay low. Raise the viewpoint change and the scores drop, so a strict threshold starts to miss real revisits. Turn on "Two places look almost the same" and a different place scores as high as the true one. That is perceptual aliasing, like two identical office corridors, and accepting it would add a wrong loop closure that can wreck the whole map. Systems guard against it by checking geometry or requiring several matches in a row.