Real-time loop detection with bags of binary words

We present a method for detecting revisited places in a image sequence in real time by using efficient features. We introduce three important novelties to the bag-of-words plus geometrical checking approach. We use FAST keypoints and BRIEF descriptors, which are binary and very fast to compute (less that 20µs per point). To perform image comparisons, we make use of a bag of words that discretises the binary descriptor space and an inverse index. We also introduce the use of a direct index to take advantage of the bag of words to obtain correspondence points between two images efficiently, avoiding a matching of complexity Θ(n²). To detect loop closure candidates, we propose managing matches in groups to increase the reliability of the candidates returned by the bag of words. We present results in three real and public datasets, with 0.7–1.7 Km long trajectories. We obtain high precision and recall rates, spending 16 ms on average per image for the feature computation and the whole loop detection process in sequences with 19000 images, one order of magnitude less than other similar techniques.