Hierarchical Strategy for Sequence-based Visual Place Recognition

Compared with a single image, image sequence contains richer and more comprehensive information on a place. In this paper, we present a sequential place recognition method based on hierarchical strategy to realize viewpoint-invariant and appearance-invariant place recognition. We first design a first-to-end match to obtain candidate sequences from the reference database for query sequence so as to narrow the range of recognition. Then, we perform 1-D convolutions and full stride pooling on the image sequence to obtain sequential descriptors. And we exploit sequential descriptors to realize sequence place recognition for query sequence on the candidate sequence. We perform extensive experiments on three benchmark VPR datasets. The results show that our approach can reduce the computational pressure effectively and has the ability to cope with severe appearance variations and viewpoint variations, while maintaining a good time performance. Meanwhile, our approach can still perform well even if the viewpoint changes extremely.