This open source MATLAB algorith presents a straightforward probabilistic appearance-based Loop Closure Detection (LCD) framework which relies on an image-to-map voting scheme based on an incremental version of Bag-of-Words methods avoiding any pre-trained technique. Feature tracking is performed using Kanade-Lucas-Tomasi (KLT) point tracker and a guided-feature-detection technique. For each tracked feature, a Tracked Word (TW) is generated by averaging the instances of the corresponding descriptors. TWs are assigned to the map representing specific locations along the trajectory, while a Bag-of-Tracked-Words (BoTW) is constructed during the navigation. Working with scale and rotation invariant SURF local-features provides a built-in robustness towards view-point and velocity variations. At query time, local-feature-descriptors vote using a k-Nearest Neighbor technique, while a binomial Probability Density Function (PDF) is adapted as a belief generator among the candidate loop closing pairs.
The Bag-of-Track-Words approach is a research code.
Bag-of-Track-Words is distributed under the terms of the MIT License.
The details of the algorithm are explained in the following publication:
Probabilistic Appearance-Based Place Recognition Through Bag of Tracked Words
Konstantinos A. Tsintotas, Loukas Bampis, and Antonios Gasteratos
IEEE Robotics and Automation Letters.
Presented at the IEEE international conference on robotics and automation (ICRA).
If you use this code, please cite:
@article{tsintotas2019botw,
title={Probabilistic Appearance-Based Place Recognition Through Bag of Tracked Words},
author={K. A. Tsintotas and L. Bampis and A. Gasteratos},
journal={IEEE Robotics and Automation Letters},
volume={4},
number={2},
pages={1737-1744},
year={2019},
month={April},
doi={10.1109/LRA.2019.2897151}
}
If you have problems or questions using this code, please contact the author (e-mail address: ktsintot@pme.duth.gr, ktsintotas@icloud.com). Contributions are totally welcome.