Shuai Yang, Wei Mou and Han Wang
Place recognition has been intensively studied in the context of robot vision. BoW-based approach gains its popularity for its efficiency and robustness using features extracted from images. Many features have been examined in the past for place recognition purpose. However, there is no such feature that can outperform others in all environments. Each feature has its own advantage, thus, they should be carefully chosen depending on the context and environments. In this paper, we propose a modified vocabulary tree with the ability of merging multiple kinds of features such that it allows users to customize different combination of features for better place recognition performance. The system is tested in real-time on real-world datasets and the experiments demonstrates the advantage of our system compared to existing approaches.
Поделиться этой статьей