Visual Place Recognition Under a High-Dimensional Subspace Clustering Perspective

• Adriel F. L. A. Da Silva UFPE
• Marcondes R. Da Silva Júnior UFPE
• Aluizio F. R. Araújo UFPE
• Adrien Durand-Petiteville UFPE

Resumo

This work deals with the topological mapping and localization problem using vision, i.e., visual place recognition. the design of a visual place recognition system is proposed, relying on a set of Self-Organizing Maps coupled with the VGG-16 feature extractor. First, the use of VGG-16 allows for extracting visual features in changing environments in a robust fashion. Next, several Self-Organizing Maps are used to map the environment by dividing the data set into categories, and then into subcategories. This approach is intended to reduce the number of images tested during the localization process, thus reducing the processing time. The proposed approach is evaluated on the St. Lucia data set and compared with two state-of-the-art methods: SeqSLAM and MPF. It presents Precision-Recall performance slightly below MPF, but better than SeqSLAM, while requiring a computation time up to 20 times smaller than MPF.

Palavras-chave: Self-organizing Maps, Visual Place Recognition, Clustering